identifying network ties from panel data: theory and …...identifying network ties from panel data:...
TRANSCRIPT
Identifying network ties from panel data: theory and an application to tax competition
Áureo de PaulaImran RasulPedro CL Souza
The Institute for Fiscal Studies
Department of Economics,
UCL
cemmap working paper CWP55/19
Identifying Network Ties from Panel Data:
Theory and an Application to Tax Competition¤
Áureo de Paula Imran Rasul Pedro CL Souzay.
October 2019
Abstract
Social interactions determine many economic behaviors, but information on social ties
does not exist in most publicly available and widely used datasets. We present results on the
identi…cation of social networks from observational panel data that contains no information on
social ties between agents. In the context of a canonical social interactions model, we provide
su¢cient conditions under which the social interactions matrix, endogenous and exogenous
social e¤ect parameters are all globally identi…ed. While this result is relevant across di¤erent
estimation strategies, we then describe how high-dimensional estimation techniques can be
used to estimate the interactions model based on the Adaptive Elastic Net GMM method. We
employ the method to study tax competition across US states. We …nd the identi…ed social
interactions matrix implies tax competition di¤ers markedly from the common assumption
of competition between geographically neighboring states, providing further insights for the
long-standing debate on the relative roles of factor mobility and yardstick competition in
driving tax setting behavior across states. Most broadly, our identi…cation and application
show the analysis of social interactions can be extended to economic realms where no network
data exists. JEL Codes: C31, D85, H71.
¤We gratefully acknowledge …nancial support from the ESRC through the Centre for the Microeconomic Ana-lysis of Public Policy (RES-544-28-0001), the Centre for Microdata Methods and Practice (RES-589-28-0001) andthe Large Research Grant ES/P008909/1 and from the ERC (SG338187). We thank Edo Airoldi, Luis Alvarez,Oriana Bandiera, Larry Blume, Yann Bramoullé, Stephane Bonhomme, Vasco Carvalho, Gary Chamberlain, An-drew Chesher, Christian Dustmann, Sérgio Firpo, Jean-Pierre Florens, Eric Gautier, Giacomo de Giorgi, MatthewGentzkow, Stefan Hoderlein, Bo Honorï¿œ, Matt Jackson, Dale Jorgensen, Christian Julliard, Maximilian Kasy,Miles Kimball, Thibaut Lamadon, Simon Sokbae Lee, Arthur Lewbel, Tong Li, Xiadong Liu, Elena Manresa, CharlesManski, Marcelo Medeiros, Angelo Mele, Francesca Molinari, Pepe Montiel, Andrea Moro, Whitney Newey, ArielPakes, Eleonora Pattachini, Michele Pelizzari, Martin Pesendorfer, Christiern Rose, Adam Rosen, Bernard Salanie,Olivier Scaillet, Sebastien Siegloch, Pasquale Schiraldi, Tymon Sloczynski, Kevin Song, John Sutton, Adam Szeidl,Thiago Tachibana, Elie Tamer, and seminar and conference participants for valuable comments. We also thankTim Besley and Anne Case for comments and sharing data. Previous versions of this paper were circulated as“Recovering Social Networks from Panel Data: Identi…cation, Simulations and an Application.” All errors remainour own. Simulation and estimation codes are available upon request.
yde Paula: University College London, CeMMAP and IFS, [email protected]; Rasul: University College Londonand IFS, [email protected]; Souza: Warwick University, [email protected]
1
1 Introduction
In many economic environments, behavior is shaped by social interactions between agents. In in-
dividual decision problems, social interactions have been key to understanding outcomes as diverse
as educational test scores, the demand for …nancial assets, and technology adoption (Sacerdote,
2001; Bursztyn et al., 2014; Conley and Udry, 2010). In macroeconomics, the structure of …rm’s
production and credit networks propagate shocks, or help …rms to learn (Acemoglu et al., 2012;
Chaney, 2014). In political economy, ties between jurisdictions are key to understanding tax setting
behavior (Tiebout, 1956; Shleifer, 1985; Besley and Case, 1994).
Underpinning all these bodies of research is some measurement of the underlying social ties
between agents. However, information on social ties does not exist in most publicly available and
widely used datasets. To overcome this limitation, studies of social interaction either postulate ties
based on common observables or homophily, or elicit data on networks. However, it is increasingly
recognized that postulated and elicited networks remain imperfect solutions to the fundamental
problem of missing data on social ties, because of econometric concerns that arise with either
method, or simply because of the cost of collecting network data.1
Two consequences are that: (i) classes of problems in which social interactions occur are un-
derstudied, because social networks data is missing or too costly to collect; (ii) there is no way
to validate social interactions analysis in contexts where ties are postulated. In this paper we
tackle this challenge by deriving su¢cient conditions under which global identi…cation of the en-
tire structure of social networks is obtained, using only observational panel data that itself contains
no information on network ties. Our identi…cation results allow the study of social interactions
without data on social networks, and the validation of structures of social interaction where social
ties have hitherto been postulated.
A researcher is assumed to have panel data on individuals = 1 for instances = 1 .
An instance refers to a speci…c observation for and need not correspond to a time period (for
example if refers to a …rm, could refer to market ). The outcome of interest for individual in
instance is and is generated according to a canonical structural model of social interactions:2
= 0
X
=1
0 + 0 + 0
X
=1
0 + + + (1)
1As detailed in de Paula (2017), elicited networks are often self-reported, and can introduce error for the outcomeof interest. Network data can be censored if only a limited number of links can feasibly be reported. Incompletesurvey coverage of nodes in a network may lead to biased aggregate network statistics. Chandrasekhar and Lewis(2016) show that even when nodes are randomly sampled from a network, partial sampling leads to non-classicalmeasurement error, and biased estimation. Collecting social network data is also a time and resource intensiveprocess. In response to these concerns, a nascent strand of literature explores cost-e¤ective alternatives to fullelicitation to recover aggregate network statistics (Breza et al., 2017).
2Blume et al. (2015) present micro-foundations based on non-cooperative games of incomplete information forindividual choice problems, that result in this estimating equation for a class of social interaction models.
2
Outcome depends on the outcome of other individuals to whom is socially tied, , and
includes characteristics of those individuals (or lagged values of ). 0 measures how the
outcome and characteristics of causally impact the outcome for . As outcomes for all individuals
obey equations analogous to (1), the system of equations can be written in matrix notation where
the structure of interactions is captured by the adjacency matrix, denoted 0. Our approach
allows for unobserved heterogeneity across individuals and common shocks to all individuals
. This framework encompasses the classic linear-in-means speci…cation of Manski (1993). In
his terminology, 0 and 0 capture endogenous and exogenous social e¤ects, and captures
correlated e¤ects. The distinction between endogenous and exogenous peer e¤ects is critical, as
only the former generates social multiplier e¤ects.
Manski’s seminal contribution set out the re‡ection problem of separately identifying endogen-
ous, exogenous and correlated e¤ects in linear models. However, it has been somewhat overlooked
that he also set out another challenge on the identi…cation of the social network in the …rst place.3
This is the problem we tackle and so expand the scope of identi…cation beyond 0, 0 and 0. Our
point of departure from much of the literature is to therefore presume 0 is entirely unknown to the
researcher. We derive su¢cient conditions under which all the entries in 0, and the endogenous
and exogenous social e¤ect parameters, 0 and 0 are globally identi…ed. By identifying the social
interactions matrix 0, our results allow the recovery of aggregate network characteristics such
as the degree distribution and patterns of homophily, as well as node-level statistics such as the
strength of social interactions between nodes, and the centrality of nodes. This is useful because
such aggregate and node-level statistics often map back to underlying models of social interaction
(Ballester et al., 2006; Jackson et al., 2017; de Paula, 2017).
The mathematical strategy for our identi…cation result is new and fundamentally di¤erent from
those employed elsewhere in this nascent literature (and does not rely on requirements on network
sparsity). However it delivers su¢cient conditions that are mild, and relate to existing results on
the identi…cation of social e¤ects parameters when 0 is known (Bramoullé et al., 2009; De Giorgi
et al., 2010; Blume et al., 2015). Our identi…cation result is also useful in other estimation contexts,
such as when a researcher has partial knowledge of 0, or in navigating between priors on reduced-
form (later denoted ¦) and structural (later denoted ) parameters in a Bayesian framework, thus
avoiding issues raised in Kline and Tamer (2016).
Global identi…cation is a necessary requirement for consistency of extremum estimators such as
those based on GMM (Hansen 1982; Newey and McFadden 1994). Our identi…cation analysis thus
3Manski (1993) highlights di¢culties (and potential restrictions) for identifying 0 0 and 0 when all individualsinteract with each other, and when this is observed by the researcher. In (1), this corresponds to 0 = ¡1, for = 1 . At the same time, he states (p. 536), “I have presumed that researchers know how individuals formreference groups and that individuals correctly perceive the mean outcomes experienced by their supposed referencegroups. There is substantial reason to question these assumptions (...) If researchers do not know how individualsform reference groups and perceive reference-group outcomes, then it is reasonable to ask whether observed behaviorcan be used to infer these unknowns (...) The conclusion to be drawn is that informed speci…cation of referencegroups is a necessary prelude to analysis of social e¤ects.”
3
provides primitives for this important condition. To estimate the model, we employ the Adaptive
Elastic Net GMM method (Caner and Zhang, 2014) because this allows us to deal with a potentially
high-dimensional parameter vector (in comparison to the time dimension in the data) including
all the entries of the social interactions matrix 0, though other estimation protocols may also be
entertained (e.g., using Bayesian methods or a priori information as previously alluded).4
We showcase the method using Monte Carlo simulations based on stylized random network
structures as well as real world networks. In each case, we take a …xed network structure 0,
and simulate panel data as if the data generating process were given by (1). We then apply the
method on the simulated panel data to recover estimates of all elements in 0, as well as the
endogenous and exogenous social e¤ect parameters (0, 0). The networks considered vary in
size, complexity, and their aggregate and node-level features. Despite this heterogeneity, we …nd
the method to perform well in all simulations. In a reasonable dimension of panel data and
with varying node numbers across simulations (), we …nd the true network structure 0 is well
recovered. For each simulated network, the majority of true links are correctly identi…ed even for
= 5, and the proportion of true non-links (zeroes in 0) captured correctly as zeros is over 85%
even when = 5. Both proportions rapidly increase with . A fortiori, we estimate aggregate
and node-level statistics of each network, demonstrating the accurate recovery of key players in
networks for example. Furthermore, biases in the estimation of endogenous and exogenous e¤ects
parameters ( ) fall quickly with and are close to zero for large sample sizes.
In the …nal part of our analysis, we apply the method to shed new light on a classic real
world social interactions problem: tax competition between US states. The literatures in political
economy and public economics have long recognized the behavior of state governors might be
in‡uenced by decisions made in ‘neighboring’ states. The typical empirical approach has been to
postulate the relevant neighbors as being geographically contiguous states. Our approach allows
us to infer the set of economic neighbors determining social interactions in tax setting behavior
from panel data on outcomes and covariates alone. In this application, the panel data dimensions
cover mainland US states, = 48, for years 1962-2015, = 53.
We …nd the identi…ed network structure of tax competition to di¤er markedly from the common
assumption of competition between geographic neighbors. The identi…ed network has fewer edges
than the geography-based network, that gets re‡ected in the far lower clustering coe¢cient in the
identi…ed network than in the geographic network (026 versus 194). With the recovered social
4The elastic net was introduced by Zou and Hastie, 2005 in part to circumvent di¢culties faced by alternativeestimation protocols (e.g., LASSO) when the number of parameters, , exceeds the number of observations, (where and follow the notation in that paper). Whereas the theoretical results on the large sample properties ofelastic net estimators usually have not exploited sparsity, several articles have demonstrated its performance in datascenarios where this occurs. For example, Zou and Hastie, 2005 consider an application to leukemia classi…cationwhere = 7 129 and = 72 (see their Section 6) and Zou and Zhang, 2009 explore a scenario where = 1 000and = 200. The favourable performance of the elastic net in these cases also relates to the literature on the‘e¤ective number of parameters’ (or ‘e¤ective degrees of freedom’) in the estimation of sparse models (Tibshiraniand Taylor, 2012). In Section 3 we provide an informal calculation for the minimum number of time periods suchthat penalized estimation is feasible in our context.
4
interactions matrix we establish, beyond geography, what covariates correlate to the existence of
ties between states and the strength of those ties. We identify non-adjacent states that in‡uence
tax setting and, more broadly, we establish that social interactions are highly asymmetric: some
states – such as Delaware, a well known low-tax state – are especially focal in driving tax setting
in other jurisdictions. We use all these results to shed new light on the main hypotheses for social
interactions in tax setting: factor mobility and yardstick competition (Tiebout, 1956; Shleifer,
1985; Besley and Case, 1994).
Our paper contributes to the literature on the identi…cation of social interactions models.
The …rst generation of papers studied the case where 0 is known, so only the endogenous and
exogenous social e¤ects parameters need to be identi…ed. It is now established that if the known
0 di¤ers from the linear-in-means example, 0 and 0 can be identi…ed (Bramoullé et al., 2009;
De Giorgi et al., 2010). Intuitively, identi…cation in those cases can use peers-of-peers, that are
not necessarily connected to individual and can be used to leverage variation from exclusion
restrictions in (1), or can use groups of di¤erent sizes within which all individuals interact among
each other (Lee, 2007). Bramoullé et al. (2009) show these conditions are met if 0 and 20 are
linearly independent, which is shown to hold generically by Blume et al. (2015). However, as made
precise in Section 2, the linear algebraic arguments employed in Bramoullé et al. (2009) or Blume
et al. (2015) do not apply when 0 is unobserved and other arguments have to be used instead.5
Our paper builds on these papers by studying the problem where 0 is entirely unknown
to the researcher. In so doing, we open up the study of social interactions to the many realms
where complete social network data does not actually exist. Closely related to our work, Blume
et al. (2015) investigate the case when 0 is partially observed. Speci…cally, Blume et al. (2015,
Theorem 6) show that if two individuals are known to not be directly connected, the parameters
of interest in a model related to (1) can be identi…ed. An alternative approach is taken in Blume
et al. (2011, Theorem 7): they suggest a parameterization of 0 according to a pre-speci…ed
distance between nodes. We do not impose such restrictions, but note that partial observability
of 0 (as in Blume et al., 2015) or placing additional structure on 0 (as in Blume et al., 2011)
is complementary to our approach as it reduces the number of parameters in 0 to be retrieved.
Bonaldi et al. (2015) and Manresa (2016) estimate models like (1) when 0 is not observed,
but where 0 is restricted to be zero so there are no endogenous social e¤ects. They use sparsity-
inducing methods from the statistics literature, but the presence of 0 in our case complicates
identi…cation non-trivially because it introduces issues of simultaneity that we address.
Rose (2015) also presents related identi…cation results for linear models like (1). Assuming
sparsity of the neighborhood structure, Rose (2015) o¤ers identi…cation conditions under rank re-
strictions on sub-matrices of the reduced form coe¢cient matrix from a regression of outcomes ()
5Alternative identi…cation approaches when 0 is known focus on higher moments (variances and covariancesacross individuals) of outcomes (de Paula, 2017), and rely on additional restrictions on the higher moments of .We also note that (1) is a spatial autoregressive model. There, 0 is also typically assumed known. Anselin (2010)reviews this literature.
5
on covariates (). Intuitively, given two observationally equivalent systems, sparsity guarantees
the existence of pairs that are not connected in either. Since observationally equivalent systems are
linked via the reduced-form coe¢cient matrix, this pair allows one to identify certain parameters in
the model. Having identi…ed those parameters, Rose (2015) shows that one can proceed to identify
other aspects of the structure (see also Gautier and Rose, 2016). This is related to the ideas in
Blume et al. (2015, Theorem 6), who show identi…cation results can be leveraged if individuals are
known not to be connected. Our main identi…cation results presented in the next section do not
rely on properties of sparse networks, and make use of plausible and intuitive conditions, whereas
the auxiliary rank conditions necessary in Rose (2015) may be computationally complex to verify.
More recently, Lewbel et al. (2019) propose an estimation strategy for the parameters 0, 0 and
0 of model (1) in the absence of network links if many di¤erent groups are able to be observed.
Finally, in the statistics literature, Lam and Souza (2019) study the penalized estimation of
model (1) when 0 is not observed, assuming the model and social interactions are identi…ed.
The statistical literature on graphical models has investigated the estimation of neighborhoods
de…ned by the covariance structure of the random variables at hand (Meinshausen and Buhlmann,
2006). This corresponds to a model where = ( ¡ 00)¡1 is jointly normal (abstracting
from covariates). On a graph with nodes corresponding to the variables in the model, an
edge between two nodes (variables) and is absent when these two variables are conditionally
independent given the other nodes. In this Gaussian model, this corresponds to a zero entry in
the inverse covariance matrix for (see, e.g., Yuan and Lin, 2007, p. 19). In the model above,
the inverse covariance matrix is ( ¡ 00)>§¡1 ( ¡ 00), where § is the variance covariance
structure for . The discovery of zero entries in this matrix is not equivalent to the identi…cation
of 0 as we study, and involves § (as do identi…cation strategies using higher moments when 0
is known).6 Related studies in the statistics literature also focus on higher moments and de…ne
neighborhoods di¤erently (Diebold and Yilmaz, 2015; Rothenhäusler et al., 2015).
Our conclusions discuss how our approach can be modi…ed, and assumptions weakened, to
integrate in partial knowledge of 0. We also discuss the next steps required to simultaneously
identify models of network formation and the structure of social interactions.
The paper is organized as follows. Section 2 presents our core result: the su¢cient conditions
under which the social interactions matrix, endogenous and exogenous social e¤ects are globally
identi…ed. Section 3 describes the high-dimensional estimation techniques used, based on the
Adaptive Elastic Net GMM method and presents simulation results from stylized and real-world
networks. Section 4 applies our methods to study tax competition between US states. Section 5
concludes. The Appendix provides proofs and further details on estimation and simulations.
6In fact, Meinshausen and Buhlmann (2006)’s neighborhood estimates (as also Lam and Souza (2019)’s) rely on(penalized) regressions of on 1 ¡1 +1 , which do not address the econometric endogeneityin estimating 0.
6
2 Identi…cation
2.1 Setup
Consider a researcher with panel data covering = 1 individuals repeatedly observed over
= 1 instances. We consider that the number of individuals in the network is …xed, but
potentially large. The aim is to use this data to identify a social interactions model, with no data on
actual social ties being available. For expositional ease, we …rst consider identi…cation in a simpler
version of the canonical model in (1), where we drop individual-speci…c () and time-constant
…xed e¤ects (), and assume is a one-dimensional regressor for individual and instance . Of
course, we later extend the analysis to include individual-speci…c and time-constant …xed e¤ects,
and also allow for multidimensional covariates , = 1 . We adopt the subscript “0” to
denote parameters generating the data, and non-subscripted parameters are generic values in the
parameter space:
= 0
X
=1
0 + 0 + 0
X
=1
0 + (2)
As outcomes for all individuals = 1 obey equations analogous to (2), the system of
equations can be more compactly written in matrix notation as:
= 00 + 0 + 00 + (3)
The vector of outcomes = (1 )0 assembles the individual outcomes in instance ; the
vector does the same with individual characteristics. , and have dimension £ 1, the
social interactions matrix 0 is £ , and 0, 0, and 0 are scalar parameters. We do not
make any distributional assumptions on beyond E(j) = 0 (or E(j) = 0 for an appropriate
instrumental variable if is also endogenous). We assume the network structure is predeter-
mined and constant, and that the number of individuals is …xed. The network structure 0 is
a parameter to be identi…ed and estimated.7
A regression of outcomes on covariates corresponds, then, to the reduced form for (3),
= ¦0 + (4)
with ¦0 = ( ¡ 00)¡1(0 + 00) and ´ ( ¡ 00)
¡1. If 0 is observed, Bramoullé
et al. (2009) note that a structure ( ) that is observationally equivalent to (0 0 0) is such
that ( ¡ 00)¡1(0 + 00) = ( ¡ 0)
¡1( + 0). This equation can be written as a
7A related set of papers instead focuses on the distribution of networks generating the pattern in data and aimsto estimate aggregate network e¤ects. Souza (2014) o¤ers several identi…cation and estimation results in this spirit.In particular, he infers the network distribution within a certain class of statistical network formation models fromoutcome data from many groups, such as classrooms, in few time periods. We instead concentrate on estimatingthe set of links for one group of size followed over = 1 instances.
7
linear equation in 0 and 20 and identi…cation is established if those matrices are linearly
independent. If 0 is not observed, the putative unobserved structure now comprises and
an observationally equivalent parameter vector will instead satisfy ( ¡ 00)¡1(0 + 00) =
(¡ )¡1(+ ). Following the strategy in Bramoullé et al. (2009) would lead to an equation
in 0 and 0, and the insights obtained in that paper then do not carry over for the case
we study when 0 is unknown.
We establish identi…cation of the structural parameters of the model, including the social
interactions matrix 0, from the coe¢cients matrix ¦0. Without data on the network 0, we treat
it as an additional parameter in an otherwise standard model relating outcomes and covariates.
Our identi…cation strategy relies on how changes in covariates reverberate through the system
and impact , as well as outcomes for other individuals. These are summarized by the entries
of the coe¢cient matrix ¦0, which, in turn, encode information about 0 and (0 0 0). A
non-zero partial e¤ect of indicates the existence of direct or indirect links between and
. When 0 = 0 (and ¦0 = 0 + 00), only direct links would produce such a correlation.
When 6= 0, both direct and indirect connections may generate a non-zero response but distant
connections will lead to a lower response. Our results formally determine su¢cient conditions to
precisely disentangle these forces.
We …rst set out …ve assumptions underpinning our main identi…cation results. Three of these
are entirely standard in the social interactions. A fourth is a normalization required to separately
identify (0, 0) from 0, and the …fth is closely related to known results on the identi…cation
of (0, 0) when 0 is known (Bramoullé et al., 2009). These Assumptions (A1-A5) deliver an
identi…ed set of up to two points.
Our …rst assumption explicitly states that no individual a¤ects himself and is a standard
condition in social interaction models:
(A1) (0) = 0, = 1 .
With Assumption (A1), we can omit elements on the diagonal of 0 from the parameter space.
We thus can denote a generic parameter vector as = (12 ¡1 )0 2 R, where
= ( ¡ 1) + 3, and is the ( )-th element of . Reduced-form parameters can be tied
back to the structural model (3) by letting ¦ : R ! R2de…ne the relation between structural
and reduced-form parameters:
¦() = ( ¡ )¡1 ( + )
where 2 R, and ¦0 ´ ¦(0).
As (and, consequently, ) is mean-independent from , E[j] = 0, the matrix ¦0 can
be identi…ed as the linear projection of on . We do not impose additional distributional
assumptions on the disturbance term, except for conditions that allow us to identify the reduced-
8
form parameters in (4). If is endogenous, i.e. E[j] 6= 0, a vector of instrumental variables
may still be used to identify ¦0. In either case, identi…cation of ¦0 requires variation of the
regressor across individuals and through instances . In other words, either E[0] (if exogeneity
holds) or E[0] (otherwise) are full-rank.
Our next assumption controls the propagation of shocks and guarantees they die as they rever-
berate through the network. This provides adequate stability in the system, and is closely related
to the concept of stationarity in network models. It implies the maximum eigenvalue norm of 00
is less than one. It also ensures (¡00) is a non-singular matrix, and so the variance of exists,
the transformation¦(0) is well-de…ned, and the Neumann expansion (¡00)¡1 =
P1=0(00)
is appropriate.
(A2)P
=1 j0(0)j 1 for every = 1 , k0k for some positive 2 R and j0j 1.
We next assume that network e¤ects do not cancel out, another standard assumption.
(A3) 00 + 0 6= 0.
The need for this assumption can be shown by expanding the expression for ¦(0), which is possible
by (A2):
¦(0) = 0 + (00 + 0)1X
=1
¡10 0 (5)
If Assumption (A3) were violated, 00 + 0 = 0 and ¦0 = 0 so the endogenous and exogenous
e¤ects balance each other out, and network e¤ects are altogether eliminated in the reduced form.8
Identi…cation of the social e¤ects parameters (0 0) requires that at least one row of 0 adds
to a …xed and known number. Otherwise, 0 and 0 cannot be separately identi…ed from 0.
Clearly, no such condition would be required if 0 was observed.
(A4) There is an such thatP
=1 (0) = 1.
Letting ´ 00 and ´ 00 denote the matrices that summarize the in‡uence of peers’
outcomes (the endogenous social e¤ects) and characteristics on one’s outcome (the exogenous social
e¤ects), respectively, the assumption above can be seen as a normalization. In this case, 0 and 0represent the row-sum for individual in and , respectively.9 In line with the literature, we
8One important case is when networks do not determine outcomes, which we interpret as 0 = 0 = 0 or with 0
representing the empty network. From equation (5), it is clear that if ¦(0) is not diagonal with constant entries,then it must be that (00 + 0) 6= 0, which implies that 0 6= 0 or 0 6= 0, and also that 0 is non-empty. Takentogether, this suggests that the observation that ¦(0) is not diagonal is su¢cient to ensure that network e¤ectsare present and Assumption (A3) is not violated.
9An alternative to Assumption (A4) is to impose the normalization on the parameters. For example, one couldnormalize ¤ = 1 and allow the network to be rescaled accordingly. In this case, ¤ = 00 would be identi…edinstead. Also = 0
0 ¤ so 0 would be identi…ed relative to 0. and are unchanged.
9
maintain that the same 0 governs the structure of both endogenous () and exogenous ()
e¤ects. We later discuss relaxing this assumption when more than one regressor is used.
Our …nal assumption provides for a speci…c kind of network asymmetry. We require the diagonal
of 20 not to be constant as one of our su¢cient conditions for identi…cation.
(A5) There exists such that ( 20 ) 6= ( 2
0 ), i.e. the diagonal of 20 is not proportional to
, where is the £ 1 vector of ones.
In unweighted networks, the diagonal of the square of the social interactions matrix captures
the number of reciprocated links for each individual or, in the case of undirected networks, the
popularity of those individuals. Assumption (A5) hence intuitively suggests di¤erential popularity
across individuals in the social network.
This assumption is related to the network asymmetry condition proposed elsewhere, such as in
Bramoullé et al. (2009). They show that when 0 is known, the structural model (2) is identi…ed
if , 0, and 20 are linearly independent. Given the remaining assumptions, this condition is
satis…ed if (A5) is satis…ed, but the converse is not true: one can construct examples in which ,
0, and 20 are linearly independent when 2
0 has a constant diagonal, so that ¦0 does not pin
down 0. The strengthening of this hypothesis is the formal price to pay for the social interactions
matrix 0 being unknown to the researcher.10
Before proceeding to our formal results, we provide a very simple illustration to shed light
on how the assumptions above come together to provide identi…cation. Suppose the observed
reduced-form matrix is,
¦0 =1
455
2
64
275 310 0
310 275 0
0 0 182
3
75
and that, following (A4), the …rst row is normalized to one. From the third row and column of ¦0,
we see there is no path of any length connecting the individual in row 3 to or from those in rows 1
10To see the strength of the assumption of Bramoullé et al. (2009) when 0 is known choose constants 1, 2,and 3 such that 1+20+3
20 = 0 Focusing on diagonal elements of this condition, we see that if the diagonal
of 20 is not proportional to the diagonal of , then 1 = 3 = 0 because diag(0) = 0. It follows that 2 = 0 if at
least one (o¤-diagonal) element of 0 is non-zero. However, the converse is not true, so that if Assumptions A1-A5do not hold, one can construct examples where ¦0 does not pin down 0. Take, for instance, = 5 with 0 and where = 0 = 1, = 15, 0 = 05, = ¡25, 0 = 05,
0 =
2
66664
0 05 0 0 0505 0 05 0 00 05 0 05 00 0 05 0 05
05 0 0 05 0
3
77775
and =
2
66664
0 0 05 05 00 0 0 05 05
05 0 0 0 0505 05 0 0 00 05 05 0 0
3
77775
Both and 0 violate (A5) (( 2) = ( 20 ) = 05 for any ), and violates (A2). Nonetheless, 0
and 20 are linearly independent and, likewise, so are , and 2. In this case, both parameter sets produce
¦ = ( ¡ 00)¡1(0 + 00) = ( ¡ )¡1( + ). This arises even as and 0 represent very di¤erent
network structures: any pair connected under is not connected under 0 and vice-versa.
10
or 2 since her outcome is not a¤ected by their covariates and their outcomes are not a¤ected by her
covariates. In other words, individual 3 is isolated and (0)13 = (0)23 = (0)31 = (0)32 = 0.
On the other hand, individuals 1 and 2 cannot be isolated as their covariates are correlated with
the other individual’s outcome, re‡ecting (A5).11 Due to the row-sum normalization of the …rst
row, (0)12 = 1. Using (A3), it can be seen that 0 is symmetric if ¦0 is symmetric. We thus
…nd that (0)21 = 1. This and (A1) map all elements of 0, and thus,
0 =
2
64
0 1 0
1 0 0
0 0 0
3
75
As the third individual is isolated, she will be only be a¤ected by her exogenous and not by
endogenous or exogenous peer e¤ects. Hence the (3 3) element of ¦0 is equal to 0 =182455
= 4.
To …nd 0, note that ( ¡ 00)¦0 = 0 + 00. Hence focussing on the (1,1) elements of the
matrices above, we …nd that 275455
¡ 0310455= 4, implying 0 = 3 (complying with (A2)). Finally, 0
is identi…ed from entry (1 2), giving 0 =310455
¡ 3275455= 5.
2.2 Main Identi…cation Results
Under the relatively mild assumptions above, we can begin to identify parameters related to the
network. These results are then useful for our main identi…cation theorems. Let 0 denote an
eigenvalue of 0 with corresponding eigenvector 0 for = 1 . Assumptions (A2) and (A3)
allow us to identify the eigenvectors of 0 directly from the reduced form. As j0j 1:
¦00 = 00 + (00 + 0)1X
=1
¡10 0 0
=
"
0 + (00 + 0)1X
=1
¡10 0
#
0
=0 + 001¡ 00
0 (6)
The in…nite sum converges as j00j 1 by (A2). The equation above implies that 0 is also
an eigenvector of ¦0 with associated eigenvalue ¦ =0+001¡00
. The fact that eigenvectors of
¦0 are also eigenvectors of 0 has a useful implication: eigencentralities may be identi…ed from
the reduced form, even when 0 is not identi…ed. As detailed in de Paula (2017) and Jackson
et al. (2017), such eigencentralities often play an important role in empirical work as they allow a
11If on the other hand, (0) = 05 6= in violation of (A5) and all agents were connected, the model wouldnot be identi…ed.
11
mapping back to underlying models of social interaction.12
Now let £ ´ f 2 R : Assumptions (A1)-(A5) are satis…edg be the structural parameter
space of interest. Our …rst theorem establishes local identi…cation of the mapping. A parameter
point 0 is locally identi…able if there exists a neighborhood of 0 containing no other which
is observationally equivalent. Using classical results in Rothenberg (1971), we show that our
assumptions are su¢cient to ensure that the Jacobian of ¦ relative to is non-singular, which, in
turn, su¢ces to establish local identi…cation.
Theorem 1. Assume (A1)-(A5). 0 2 £ is locally identi…able.
An immediate consequence of local identi…cation is that the set f 2 £ : ¦() = ¦(0)g is
discrete (i.e. its elements are isolated points). The following corollary establishes that ¦ is a proper
function, i.e. the inverse image ¦¡1() of any compact set ½ R2is also compact (Krantz and
Parks, 2013, p. 124). Since it is discrete, the identi…ed set must be …nite.
Corollary 1. Assume (A1)-(A5). Then ¦(¢) is a proper mapping. Moreover, the set f : ¦() =
¦(0)g is …nite.
Under additional assumptions, the identi…ed set is at most a singleton in each of the partitioning
sets £¡ ´ £ \ f + 0g and £+ ´ £ \ f + 0g.13 Since £ = £¡ [ £+, if the sign
of 00 + 0 is unknown, the identi…ed set contains, at most, two elements. In the theorem that
follows, we show global identi…cation only for 2 £+, since arguments are mirrored for 2 £¡.
Theorem 2. Assume (A1)-(A5), then for every 2 £+ we have ¦() = ¦(0) ) = 0. That
is, 0 is globally identi…ed with respect to the set £+.
Similar arguments apply if Theorem 2 instead were to be restricted to 2 £¡. The proof of the
corollary below is immediate and therefore omitted.
Corollary 2. Assume (A1)-(A5). If 00 + 0 0, then the identi…ed set contains at most one
element, and similarly if 00 + 0 0. Hence, if the sign of 00 + 0 is unknown, the identi…ed
set contains, at most, two elements.14
12To identify the eigencentralities, we identify the eigenvector that corresponds to the dominant eigenvalue. If 0
is non-negative and irreducible, this is the (unique) eigenvector with strictly positive entries, by the Perron-FrobeniusTheorem for non-negative matrices (see Horn and Johnson, 2013, p.534).
13The global inversion results we use are related to, but di¤erent from, those used by Komunjer (2012), Leeand Lewbel (2013) and Chiappori et al. (2015). Those authors use variations on a classical inversion result ofHadamard. In contrast, we employ results on the cardinality of the pre-image of a function, relying on less stringentassumptions. Speci…cally, while the classical Hadamard result requires that the image of the function be simply-connected (Theorem 6.2.8 of Krantz and Parks, 2013), the results we rely on do not.
14Under some special conditions, the mirror image of 0 can be characterized from equation (5). If ¡0 sat-is…es Assumption (A4), we may set ¤ = ¡0, ¤ = 0, ¤ = ¡0 and ¤ = ¡0. Then 00 + 0 =¡(¤¤ + ¤). Also note that
P1=1
¡10
0 = ¡P1
=1(¤)¡1( ¤), and so (00 + 0)
P1=1
¡10
0 =(¤¤ + ¤)
P1=1(
¤)¡1( ¤). It follows that ¦(0) = ¦(¤), where ¤ = (¤ ¤ ¤ ¤).
12
We now turn our attention to the problem of identifying the sign of 00 + 0 from the obser-
vation of ¦0. This would then allow us to establish global identi…cation using Theorem 2. It is
apparent from (5) that if 0 0 and (0) ¸ 0, for all = f1 g the o¤-diagonal elements
of ¦0 identify the sign of 00 + 0.
Corollary 3. Assume (A1)-(A5). If 0 0 and (0) ¸ 0, the model is globally identi…ed.
Real world applications often suggest endogenous social interactions are positive (0 0), in
which case global identi…cation is fully established by Corollary 3. On the other hand, if 0 0
(which is so if outcomes are strategic substitutes, for example), 0 in (5) alternates signs with ,
and the o¤-diagonal elements no longer carry the sign of 00 + 0. Nonetheless, if 0 is non-
negative and irreducible (i.e., not permutable into a block-triangular matrix or, equivalently, a
strongly connected social network), the model is also identi…able without further restrictions on
0:
Corollary 4. Assume (A1)-(A5), (0) ¸ 0 and 0 is irreducible. If 0 has at least two real
eigenvalues or j0j p22, then the model is globally identi…ed.
Corollary 4 holds if there are at least two real eigenvalues, or if 0 is appropriately bounded.
Since 0 is non-negative, it has at least one real eigenvalue, by the Perron-Frobenius Theorem. If
0 is symmetric, for example, its eigenvalues are all real, and Corollary 4 holds. It also holds if
(0) · 0, as we can re-write the model as 0 = ¡j0j where j0j, is the matrix whose entries
are the absolute values of the entries in 0. In any case, the bound on j0j is su¢cient and holds
in most (if not all) empirical estimates we are aware of obtained from either elicited or postulated
networks, and in our application on tax competition.
2.3 Extensions
2.3.1 Individual Fixed E¤ects
We observe outcomes for = 1 individuals repeatedly through = 1 instances.
If corresponds to time, it is natural to think of there being unobserved heterogeneity across
individuals, , to be accounted for when estimating ¦0. The structural model (2) is then,
= 0
X
=1
0 + 0 + 0
X
=1
0 + +
which can be written in matrix form as,
= 00 + 0 +00 + ¤ +
13
where ¤ is the vector of …xed e¤ects. Individual-speci…c and time-constant …xed e¤ects can be
eliminated using the standard subtraction of individual time averages. De…ning ¹ = ¡1P
=1 ,
¹ = ¡1P
=1 and ¹ = ¡1P
=1 ,
¡ ¹ = 00 ( ¡ ¹) + ( ¡ ¹)0 +0 ( ¡ ¹) 0 + ¡ ¹
if 0 is does not change with time. Identi…cation from the reduced form follows from previous
theorems, since ¦0 is unchanged when regressing ¡ ¹ on ¡ ¹.15
2.3.2 Common Shocks
We next allow for unobserved common shocks to all individuals in the network in the same instance
. Such correlated e¤ects can confound the identi…cation of social interactions. As we have not
placed any distributional assumption on the covariance matrix of the disturbance term, our analysis
readily incorporates correlated e¤ects that are orthogonal to . When this is not the case, one
possibility is to model the corrected e¤ects explicitly. The model then is,
= 00 + 0 + 00 + +
where is a scalar capturing shocks in the network common to all individuals. Let ¦01 =
( ¡ 00)¡1 and ¦02 = (0 + 00) such that ¦0 = ¦01¦02. The reduced-form model is,
= ¦0 + ¦01+
We propose a transformation to eliminate the correlated e¤ects: exclude the individual-invariant
, subtracting the mean of the variables at a given period (global di¤erencing). For this purpose,
de…ne = 10. We note that in empirical and theoretical work it is customary to strengthen
Assumption (A4) and require that all rows of 0 sum to one if no individual is isolated (see for
example Blume et al., 2015). This strengthened assumption is usually referred to as row-sum
normalization, and is stated below:
(A4’) For all = 1 we have thatP
=1(0) = 1.
This can be written compactly as 0 = . In this case, 0 can be interpreted as the normalized
adjacency matrix. Under row-sum normalization we have that,
( ¡) = ( ¡) ( ¡ 00)¡1 (0 + 00) + ( ¡) ( ¡ 00)
¡1
= ( ¡)¦0 + ( ¡ )
15As is the case in panel data, this would require strict exogenity (E[j] = 0 for any and ) or predeterminederrors (E[j] = 0 for ¸ ) so that the matrix ¦0 can be consistently estimated.
14
because ( ¡) ( ¡ 00)¡1 = 0 if Assumption (A4’) holds. It then follows that ~¦0 =
( ¡)¦0 is identi…ed. The next proposition shows that, under row-sum normalization of 0, ¦0
is identi…ed from ~¦0 (and, as a consequence, the previous results immediately apply).
Proposition 1. If 0 is diagonalizable and row-sum normalized, ¦0 is identi…ed from ~¦0.
Under row-sum normalization of 0, a common group-level shock a¤ects individuals homo-
geneously since ( ¡ 00)¡1 = ( + 00 + 20
20 + ¢ ¢ ¢ ) =
1¡0, which is a vector with
no variation across entries. Consequently, global di¤erencing eliminates correlated e¤ects and
( ¡ ) ( ¡ 00)¡1 = ( ¡ 00)
¡1 ( ¡) = 0. In the absence of row-sum normaliza-
tion, global di¤erencing does not ensure that correlated e¤ects are eliminated. To see this, note
that ( ¡ 00)¡1 is no longer row-sum normalized and, crucially, ( ¡ 00)
¡1 is not a vector
with constant entries.
The next proposition makes this point formally, that the stronger Assumption (A4’) is necessary
to eliminate group-level shocks, by showing it is not possible to construct a data transformation
that eliminates group e¤ects in the absence of row-sum normalization.
Proposition 2. De…ne 0 = ( ¡ 00)¡1. If in space £ = f 2 R : Assumptions (A1)-(A5)
are satis…edg there are matrices (1)0
()0 such that [
(1)0
¢ ¢ ¢
()0] has rank , then
the only transformation such that ( ¡ ~)( ¡ 00)¡1 = 0 is ~ = .
It is useful to be able to test for row-sum normalization (A4’) as it enables common shocks to
be accounted for in the social interactions model. This is possible as,
¦0 = 0+ (00 + 0)1X
=1
¡10 0
=
"
0 + (00 + 0)1X
=1
¡10
#
=0 + 01¡ 0
(7)
The last equality follows from the observation that, under row-normalization of 0, 0 = 0 =
, 0. This implies ¦0 has constant row-sums, which suggests row-sum normalization is testable.
In the Appendix we derive a Wald test statistic to do so.16
16For ease of explanation, in the Appendix we derive the test under the asymptotic distribution of the OLSestimator. The test generally holds with minor adjustments for estimators with known asymptotic distributions.
15
2.3.3 Multivariate Covariates
Next allowing for multivariate of dimension £ , the reduced-form model (4) is,
=X
=1
¦0 +
where ¦0 = ( ¡ 00)¡1 ¡
0 + 00
¢, refers to the -th column of , and 0 and
0 select the -th element of -dimensional 0 and 0, respectively. The previous identi…cation
results then apply sequentially to each ¦0, = 1 . In fact, we only then need to maintain
= 00 for one covariate. It is therefore possible to allow the structure of endogenous and
exogenous social e¤ects to di¤er for ¡ 1 of the covariates. With covariates, equation (3) is,
= 00 +X
=1
0 +X
=1
00 +
Let 0 = 0 be the case for = 1. Then, having identi…ed 0 and 0 from ¦10,
( ¡ 00)¦0 = 0 + 00
for = 2 . The parameter 0 then corresponds to the diagonal elements of ( ¡ 00)¦0
and the o¤-diagonal entries correspond to the o¤-diagonal elements of 00. If Assumption
(A4) holds for every = 1 , we can identify 0 and thus 0 for every = 1 .17
3 Implementation
We now transition from our core identi…cation results to their practical implementation. As this is
a high-dimensional estimation problem, our preferred approach makes use of the Adaptive Elastic
Net GMM (Caner and Zhang, 2014), that is based on the penalized GMM objective function. Given
the identi…cation results presented in Section 2, the populational version of the GMM objective
function will be uniquely minimized at the true parameter vector.
After setting out the estimation procedure, we showcase the method using Monte Carlo simu-
lations based both on stylized random network structures as well as real world networks. In each
case, we take a …xed network structure 0, and simulate panel data as if the data generating
process were given by (1). We apply the method on the simulated panel data to recover estimates
of all elements in 0, as well as the endogenous and exogenous social e¤ect parameters.
17Blume et al. (2015) also study the case in which the social structure mediating endogenous and exogenous sociale¤ects might di¤er. When is known and there is partial knowledge of the endogenous social interaction matrix0, they show that the parameters of the model can be identi…ed (their Theorem 6). Analogously, when there areenough unconnected nodes in each of the social interaction matrices represented by and 0, and the identityof those nodes is known, identi…cation is also (generically) possible (their Theorem 7).
16
3.1 Estimation
The parameter vector to be estimated is high-dimensional: = (12 ¡1 )0 2 R,
where = ( ¡ 1) + 3 and is the ( )-th element of the £ social interactions matrix
0. To be clear, in a network with individuals, there are ( ¡ 1) potential interactions
because individuals could interact with everyone else but herself (which would violate Assumption
A1). As a consequence, even with a modest , there are many more parameters to estimate and
is large. For example, a network with = 50 implies more than two thousand parameters
to estimate. While we consider (and thus ) is …xed, we still refer to as high-dimensional.
OLS estimation requires ¿ () ¿ ), so many more time periods than individuals: a
requirement often met in …nance data sets (van Vliet, 2018) or in other …elds (see, e.g., Section 4.2
in Rothenhäusler et al., 2015). Instead, to estimate a large number of parameters with limited data
we utilize high-dimensional estimation methods, that are the focus of a rapidly growing literature.
However, the identi…cation results presented in Section 2 apply more broadly and irrespective of
the estimation procedure.
Sparsity is a key assumption underlying all high-dimensional estimation techniques. In the
context of social interactions, we say that 0 is sparse if ~, the number of non-zero elements of
0, is such that ~ ¿ . The notion of sparsity thus depends on the number of of time periods:
although and are …xed, ~ itself can grow with . Sparsity corresponds to assuming that
individuals in‡uence or are in‡uenced by a small number of others, relative to the overall size of
the potential network and the time horizon in the data. As such, sparsity is typically not a binding
constraint in social networks analysis.18
In the estimation of sparse models, the “e¤ective number of parameters” (or “e¤ective degrees
of freedom”) relates to the number of variables with non-zero estimated coe¢cients (Tibshirani and
Taylor, 2012). In the context of the current social network model (and the Elastic Net estimator
on which the estimation strategy below builds on), this is approximately equivalent to the density
of the network times the number of parameters . We then require this number to be smaller
than . Implicitly, this calculation provides a rough assessment on the minimum required . For
example, with = 30 and a network with 2% of potential links in place, this implies should be
larger than 18.
Finally, to reiterate, our identi…cation results themselves do not depend on the sparsity of
networks. In particular, Assumptions A1 to A5 do not impose restrictions on the number of links
18For example, common stylized networks are sparse, such as: (i) star: all individuals receive spillovers from thesame individual; (ii) lattice: each individual is a source of spillover only to one other individual; (iii) interactions inpairs or triads or small groups, such as those described by De Giorgi et al. (2010); and (iv) small world networks(Watts, 1999). Prominent real world economic networks are also sparse. For example, in individual-level eliciteddata from AddHealth on teenage friendships (de…ned as reciprocated nominations), the density of links is around2% of all feasible links. In …rm-level data, the density of production networks in the US is less than 1% of all feasiblelinks (Atalay et al., 2011).
17
in 0, or ~.19
Our preferred approach estimates the interaction matrix in the reduced form while penalizing
and imposing sparsity on the structural object 0. We impose sparsity and penalization in the
structural-form matrix 0 because this is a weaker requirement than imposing sparsity and penal-
ization in the reduced-form matrix ¦0.20 To accomplish this, we make use of the Adaptive Elastic
Net GMM (Caner and Zhang, 2014), that is based on the penalized GMM objective function,
( ) ´ ()0 () + 1
X
=16=
jj+ 2
X
=16=
jj2 (8)
where = (12 ¡1 )0 with dimension = ( ¡ 1) + 3, and 1 and 2 are the
penalization terms. The term ()0 () is the unpenalized GMM objective function with
moment conditions based on the orthogonality between the structural disturbance term and the
covariates: () =P
=1 [1()0 ¢ ¢ ¢ ()
0]0, () = ¡ ( ¡ )¡1 ( + ) . There
are ´ 2 moment conditions since is orthogonal to , for each = 1 . Hence the
GMM weight matrix is of dimension 2£2, symmetric, and positive de…nite. For simplicity,
we use = 2£2. Note that if is econometrically endogenous, one can also exploit moment
conditions with respect to available instrumental variables.21
Given the identi…cation results presented in Section 2, if 6= 0 and does not belong to the
identi…ed set, then ¦() 6= ¦(0). Consequently, the populational version of the GMM objective
function is uniquely minimized at the true parameter vector 0.
The penalization terms in (8) is what makes this di¤erent from a standard GMM problem.
The …rst term, 1P
=16= jj, penalizes the sum of the absolute values of , i.e. the sum
of the strength of links, for all node-pairs. The second term, 2P
=16= jj2, penalizes the
sum of the square of the parameters. This term has been shown to provide better model-selection
properties, especially when explanatory variables are correlated (Zou and Zhang, 2009). The …rst
stage estimate is,~() = (1 + 2 ) ¢ argmin
2R ( ) (9)
where (1 + 2 ) is a bias-correction term also used by Caner and Zhang (2014).
19If ! 1, Assumption A2 would imply vanishing (0) entries. As highlighted previously, we consider tobe …xed, in line with many practical applications. Furthermore, Assumption A2 is used to represent inverse matricesas Neumann series in our identi…cation results. What is necessary for this to hold is that a sub-multiplicative normon be less than one. Here we use a speci…c norm (i.e., the maximum row sum norm), but other (induced) normsare also possible (i.e., the 2-norm or the 1-norm) (see Horn and Johnson, 2013, Chapter 5.6).
20Note that even if is sparse, ¦ may not be sparse. In Appendix B.1, we show that [¦0] = 0 if, and only if,there are no paths between and in 0, and so the pair is not connected. So sparsity in ¦0 is understood as 0
being ‘sparsely connected’, which is a stronger assumption than sparsity in 0.21For expositional ease, we describe estimation in the context of the reduced form model (4), thereby abstaining
from individual …xed or correlated e¤ects. As the GMM estimator uses moments between the structural disturbanceterms and covariates, this endogeneity is built into the estimation procedure.
18
Depending on the choice of 1, some ’s will be estimated as exact zeros. A larger share
of parameters will be estimated as zeros if 1 increases. The penalization also shrinks non-zero
estimates towards zero. A second (adaptive) step provides improvements by re-weighting the
penalization by the inverse of the …rst-step estimates (Zou, 2006):
() = (1 + 2 ) ¢ argmin2R
8>>>>><
>>>>>:
()0 () + ¤1
X
f: ~ 6=0=1
6=g
jj
j ~j+ 2
X
f: ~ 6=0=1
6=g
jj2
9>>>>>=
>>>>>;
(10)
where ~ is the ( )-th element of the …rst-step estimate of , and we follow Caner and Zhang
(2014) to set = 25. Elements ~ estimated as zeros in the …rst stage are kept as zero in the
second stage, because ~ = 0 implies the e¤ective penalization is in…nite. We write = (1 ¤1 2)
as the …nal set of penalization parameters. Conditional on , the estimate of the Adaptive Elastic
Net GMM procedure is (). Finally, we update the estimates of 0, 0 and 0 on a regression
using peers-of-peers as instruments, similar to Bramoullé et al. (2009), but using the network as
estimated in (10). This …nal step is not necessary but performs better in small samples. As in
Caner and Zhang (2014, p. 35), the penalization parameters are chosen by the BIC criterion.
This balances model …t with the number of parameters included in the model.22
In Appendix B.2 we provide further implementation details, including the choice of initial
conditions. Of course, other estimation methods are available and our identi…cation results do not
hinge on any particular estimator. Our aim is to demonstrate the practical feasibility of using
the Adaptive Elastic Net estimator, rather than claim it is the optimal estimator.23 Indeed, in
Appendix B.3 we show how OLS can also be used to estimate if is su¢ciently large. This
makes precise the bene…ts of penalized estimation for any given and highlights that sparsity is
not required for our identi…cation results.
3.2 Simulations
We showcase the method using Monte Carlo simulations based both on stylized random network
structures as well as real world networks. We describe the simulation procedures, results and
22Following Caner and Zhang, 2014, the choice of , which we denote as , is the one that minimizes
BIC() = log
·
³()
´0
³()
´¸
+³()
´¢log
where ³()
´counts the number of non-zero coe¢cients among f12 ¡1g. (See also Zou et al., 2007.)
23For example, Manresa (2016) also relies on a Lasso-related methodology but restricts 0 to be zero and soignores endogenous social e¤ects. If instrumental variables are available, Lam and Souza (2016) propose estimating(1) directly using the Adaptive Lasso and exploiting sparsity of the estimated 0. Gautier and Rose (2016) extendthe (identi…cation-robust) Self-Tuning Instrumental Variable estimator in Gautier and Tsybakov (2014).
19
robustness checks in more detail in the Appendix. Here we just provide a brief overview to
highlight how well the method works to recover social networks even in relatively short panels.
For each simulated network, we take a …xed network structure 0, and simulate panel data
as if the data generating process were given by (1). We then apply the method on the simulated
panel data to recover estimates of all elements in 0, as well as the endogenous and exogenous
social e¤ect parameters (0, 0). Our result identi…es entries in 0 and so naturally recovers links
of varying strength. It is long recognized that link strength might play an important role in social
interactions (Granovetter, 1973). Data limitations often force researchers to postulate some ties to
be weaker than others (say, based on interaction frequency). In contrast, our approach identi…es
the continuous strength of ties, 0, where 0 0 implies node in‡uences node .
The stylized networks we consider are a random network, and a political party network in which
two groups of nodes each cluster around a central node. The real world networks we consider are
the high-school friendship network in Coleman (1964) from a small high school in Illinois, and
one of the village networks elicited in Banerjee et al. (2013) from rural Karnataka, India. These
networks vary in size, complexity, and their aggregate and node-level features. All four networks
are also sparse. For the stylized networks, we …rst assess the performance of the estimator for a
…xed network size, = 30. We simulate the real-world networks using non-isolated nodes in each
(so = 70 and 65 respectively).24
Despite the heterogeneity across network scenarios, the method performs well in all simulations.
Figure A1 shows the simulation results. Each Panel presents a di¤erent metric as we vary for
each simulated network. Panel A shows that for each network, the proportion of zero entries in
0 correctly estimated as zeros is above 90% even when exploiting a small number of time periods
( = 5). The proportion approaches 100% as grows. Conversely, Panel B shows the proportion
of non-zeros entries estimated as non-zeros is also high for small . It is above 70% from = 5
for the Erdos-Renyi network, being at least 85% across networks for = 25, and increasing as
grows. As discussed above, the Adaptive Elastic Net estimator is better in recovering true zero
entries because it is a well-known feature that shrinkage estimators tend to shrink small parameters
to zero.
Panels C and D show that for each simulated network, the mean absolute deviation between
estimated and true networks for and ¦ falls quickly with and is close zero for large sample
sizes. Finally, Panels E and F show that biases in the endogenous and exogenous social e¤ects
parameters, and , also fall quickly in . The fact that biases are not zero is as expected for
small , being analogous to well-known results for autoregressive time series models.25
In the Appendix we show the robustness of the simulation results to: (i) varying network sizes
and node de…nitions in the real work network of Banerjee et al. (2013); (ii) alternative parameter
24As in Bramoullé et al., 2009, we exclude isolated nodes because they do not conform with row-sum normalization.25The bias in spatial auto-regressive models with small number of observations even when the network is observed
is similarly documented by Mizruchi and Neuman (2008), Farber et al. (2009), Smith (2009), Neuman and Mizruchi(2010), and Wang et al. (2014).
20
choices and richening up the structure of shocks across nodes. We also demonstrate the gains from
using the Adaptive Elastic Net GMM estimator over alternative estimators, such as the Adaptive
Lasso estimator and OLS.
4 Application: Tax Competition Between US States
Our identi…cation result can be used to shed new light on a classic social interactions problem:
tax competition between US states (Wilson, 1999). Since the seminal empirical studies in tax
competition between jurisdictions (Case et al., 1989; Case et al., 1993), it has been well-recognized
that de…ning competing ‘neighbors’ is the central empirical challenge, and theory cannot resolve the
issue. Two mechanisms have been argued to drive the structure of interactions across jurisdictions:
factor mobility and yardstick competition.
On factor mobility, Tiebout (1956) …rst argued that labor and capital can move in response to
di¤erential tax rates across jurisdictions. Factor mobility leads naturally to the postulated social
interactions matrix being: (i) geographic neighbors given labor mobility; and (ii) jurisdictions with
similar economic or demographic characteristics, given capital mobility (Case et al., 1989).26
A second mechanism occurs through political economy channels (Shleifer, 1985). In particular,
yardstick competition between jurisdictions is driven by voters making comparisons between states
to learn about their own politician’s quality. Besley and Case (1995) formalize the idea in a
model where voters use taxes set by governors in neighboring states to infer their own governor’s
quality. This generates informational externalities across jurisdictions, forcing incumbents into
yardstick competition, where their tax setting behavior is determined by what other incumbents
do. Yardstick competition leads naturally to the postulated interactions matrix corresponding to
a matrix of ‘political neighbors’: other states that voters make comparisons to.
This application shows the practical use of our approach to recover social interactions in a
setting in which the number of nodes and time periods is relatively low: the data covers mainland
US states, = 48, for years 1962-2015, = 53. Our approach identi…es the structure of social
interactions among ‘economic neighbors’, that we denote . We contrast this against a null
hypothesis that states are only in‡uenced by their geographic neighbors, , as postulated by
Besley and Case, 1995 and shown in Figure 1A. With recovered, we can establish, beyond
geography, what predicts the existence and strength of ties between states. Finally, relative to
, we conduct simulations using to assess the equilibrium propagation of tax setting
shocks across mainland US states. Taken together, this body of evidence allows us to provide
novel insights related to the role of factor mobility and yardstick competition in driving tax setting
behavior across US states.
26A body of evidence …nds that tax bases are mobile in response to tax di¤erentials (Hines, 1996; Devereux andGri¢th, 1998; Kleven et al., 2013, 2014)
21
4.1 Data and Empirical Speci…cation
We denote state tax liabilities for state in year as , covering state taxes collected from real per
capita income, sales and corporate taxes. We measure this using a series constructed from data
published annually in the Statistical Abstract of the United States. Our series covers mainland
states ( = 48) for years 1962-2015, ( = 53), therefore extending the sample used by Besley and
Case (1995), that runs from 1962-1988 ( = 26).27 The outcome considered, ¢ , is the change
in tax liabilities between years and (¡ 2) because it might take a governor more than a year to
implement a tax program. Their model implies a standard social interactions speci…cation for the
tax setting behavior of state governors:
¢ = X
=1
0¢ + X
=1
0 + + + + (11)
Tax setting behavior is thus determined by (i) endogenous social e¤ects arising through neighbors’
tax changes (P
=10¢ ); (ii) exogenous social e¤ects arising through the economic/demographic
characteristics of neighbors (P
=10); (iii) state ’s characteristics (), that include income
per capita, the unemployment rate, and the proportion of young and elderly. All speci…cations
include state and time e¤ects ( ), so allowing for time-invariant unobserved heterogeneity
across states, and for common (macroeconomic) shocks. Due to the inclusion of the time e¤ects
, we normalize the rows of to one. Table A7 presents descriptive statistics for the Besley
and Case (1995) sample and our extended sample.
Much of the earlier literature focuses on endogenous social e¤ects and ignores exogenous social
e¤ects by setting = 0. Our identi…cation result allows us to relax this constraint and thus
estimate the full typology of social e¤ects described by Manski (1993). This is important because
only endogenous social e¤ects lead to social multipliers, and are crucial to identify as they can
lead to a race-to-the-bottom or sub-optimal public goods provision (Brennan and Buchanan, 1980;
Wilson, 1986; Oates and Schwab, 1988).
After estimating the neighborhood matrix, we follow Besley and Case (1995) and estimate the
model instrumenting for ¢ using neighbors’ lagged change in income per capita, and neighbors’
lagged change in unemployment rate. These instruments are in the spirit of using exogenous social
e¤ects to instrument for neighbor’s tax changes. However, given our approach allows us to estimate
exogenous social e¤ects ( 6= 0), these instruments will generally be weaker when estimating the
full speci…cation in (11). We thus follow Bramoullé et al. (2009) and De Giorgi et al. (2010), and
also instrument neighbors’ tax changes with neighbor-of-neighbor characteristics.
27Besley and Case (1995) test their political agency model using a two equation set-up: (i) on gubernatorialre-election probabilities; and (ii) on tax setting. Our application focuses on the latter because this represents asocial interaction problem. They use two tax series: (i) TAXSIM data (from the NBER) which runs from 1977-88;and (ii) state tax liabilities series constructed from data published annually in the Statistical Abstract of the USthat runs from 1962-1988. All their results are robust to either series. We extend the second series.
22
4.2 Preliminary Findings
Table 1 presents our preliminary …ndings and comparison to Besley and Case (1995). Column 1
shows OLS estimates of (11) where the postulated social interactions matrix is based on geographic
neighbors, exogenous social e¤ects are ignored so = 0 and the panel includes all 48 mainland
states but runs only from 1962-1988 as in Besley and Case (1995). Social interactions in‡uence
gubernatorial tax setting behavior: b = 375. Column 2 shows this to be robust to instrument-
ing neighbors’ tax changes using the instrument set proposed by Besley and Case (1995). b2 is
more than double the magnitude of b suggesting tax setting behaviors across jurisdictions are
strategic complements, and OLS estimates are heavily downward-biased.
Columns 3 and 4 replicate both speci…cations over the longer sample period we construct. The
evidence con…rms Besley and Case’s (1995) …nding on social interactions to be robust in a longer
sample period. We again note that b2 is more than double the magnitude of b. The result
in Column 4 implies that for every dollar increase in the average tax rates among geographic
neighbors, a state increases its own taxes by 61 cents. This is similar to the headline estimate of
Besley and Case (1995).28
4.3 Endogenous and Exogenous Social Interactions ( and )
We now move beyond much of the earlier political economy and public economics literature to …rst
establish whether there are endogenous and exogenous social interactions in tax setting behavior.
We …rst focus on the endogenous and exogenous social interaction parameters, and in the next
subsection we detail the identi…ed social interactions matrix, . Column 1 of Table 2 shows
the initial estimates obtained from the Adaptive Elastic Net procedure where = 0. Columns
2 and 3 show the resulting OLS and 2SLS estimates for : b2 = 641 b = 378 0.29
Columns 4 to 6 estimate the full model in (11). Columns 5 and 6 show the OLS and 2SLS estimates
of are smaller, and less precisely estimated when exogenous social e¤ects are allowed. This is
not surprising given that the instrument set is based on neighbors’ characteristics, many of which
are directly controlled for in (11), thus reducing the e¤ective variation induced by the instrument.
Hence, in Column 7, we report 2SLS estimates based on instruments using neighbor-of-neighbor
characteristics. This represents our preferred speci…cation: b2 = 608 (with a standard error of
220). This value also meets the requirements on in Corollaries 3 and 4 for global identi…cation.
In short, there is robust evidence of endogenous social interactions in tax setting behavior of
governors across states.30
28Nor is the magnitude very di¤erent from earlier work examining …scal expenditure spillovers. For example,Case et al. (1989) …nd that US state government levels of per-capita expenditures are signi…cantly impacted bythe expenditures of their neighbors, with the size of the impact being that a one dollar increase in neighbors’expenditures leads to an increase in own-state expenditures by seventy cents.
29We report robust standard errors and so do not adjust them for the fact that is estimated.30Table A8 shows the full set of exogenous social e¤ects (so Columns 1 to 4 refer to the same speci…cations as
23
4.4 Identi…ed Social Interactions Matrix ()
Figure 1B shows how the structure of economic () and geographic networks () di¤er,
where connected edges imply that two states are linked in at least one direction (either state
causally impacts state taxes in , and/or vice versa). This comparison makes it clear whether
all states geographically adjacent to matter for its tax setting behavior and whether there are
non-adjacent states that in‡uence its tax rate.
The left-hand panel of Figure 1B shows the network of geographic neighbors (whose edges
are colored blue), onto which we have superimposed the edges that are not identi…ed as links in
; these dropped edges are indicated in red. This …rst implies that not all geographically
adjacent states are relevant for tax setting behavior. The right-hand panel of Figure 1B adds new
edges identi…ed in that are not part of . These represent non-adjacent states through
which social interactions occur. This implies the existence of spatially dispersed social interactions
between states. The implication is that for tax-setting behavior, economic distance is imperfectly
measured if we simply assume that interactions depend only on geographical distance. As detailed
below, this has many implications for the economics of tax competition.
As Table 3 summarizes, has 214 edges, while has only 144 edges. States are less
connected than implied by postulating geographic networks. and have 79 edges in
common. However, has 135 edges that are absent in . Hence, while geography remains
a key determinant of tax competition, the majority of geographical neighbors (135214 = 63%)
are not relevant for tax setting. There are 65 edges that exist only in , so although there are
fewer edges in , the identi…ed social interactions are more spatially dispersed than under the
assumption of geographic networks. This is re‡ected in the far lower clustering coe¢cient in
than in (026 versus 194).31
4.5 Strength of Ties and Reciprocity
Our estimation strategy identi…es the continuous strength of ties, 0, where 0 0 is inter-
preted as state in‡uencing outcomes in state . This is useful because recent developments in
tax competition theory, using insights from the social networks literature, suggest links need not
be reciprocal or of symmetric strength (Janeba and Osterleh, 2013).
Figure 2A shows the distribution of 0’s across edges in (conditional on 0 0).
The strength of ties between pairs of states varies greatly. The mean strength of ties is 19, that
is higher than the median strength, 085, suggesting many weak ties. At the other end of the
Columns 4 to 7 in Table 2). Exogenous social e¤ects operate through economic neighbors’ income per capita andunemployment rate. Demographic characteristics of economic neighbors to state do not impact its tax rate.
31The clustering coe¢cient is the frequency of the number of fully connected triplets over the total number oftriplets. Other metrics can also be used to provide a scalar comparison of and . One way to do so isto reshape both matrices as vectors of length (48 £ 47) and to compute their correlation. Doing so, we obtain acorrelation coe¢cient of 322.
24
distribution, the strongest 10% of ties have weight above 6.
On the reciprocity of ties, Table 3 reveals that only 292% of edges in are reciprocal (all
edges in are reciprocal by construction). Hence, tax competition is both spatially disperse
and highly asymmetric. In most cases where tax setting in state is in‡uenced by taxes in state
, the opposite is not true.
Panels B and C in Figure 2 illustrate this for California, indicating the strength of each tie
(). Figure 2B shows the in-network for California: those states causally impacting tax-
setting in California. Some geographic neighbors to California in‡uence its tax setting behavior
(Nevada and Oregon), although these ties are weak. On the other hand, non-adjacent states
in‡uence California (Colorado, Maine), and these in-network ties are stronger than geographically
adjacent in-network ties. Figure 2C shows the out-network for California, again indicating each tie
strength (): those states whose taxes are in‡uenced by taxes in California. We see that
none of the geographic neighbors to California are in‡uenced by its tax setting behavior, whereas
a number of non-adjacent states are in‡uenced (including East Coast states such as Virginia, and
Southern states, such as Louisiana). When states are in‡uenced by taxes in California, these ties
tend to be relatively strong: > 19 for all …ve in-network ties.
Given common time shocks in (11), row-sum normalization is required and ensuresP
0 =
1. Hence, for every state there will be at least one economic neighbor state ¤ impacts it, so that
0¤ 0. This just reiterates that social interactions matter. On the other hand, our procedure
imposes no restriction on the derived columns of . It could be that a state does not a¤ect any
other state. Examining this possibility directly in , we see this occurs for Minnesota, New
Jersey, New Mexico, Vermont, and Wisconsin. These states have an out-degree of zero. Their tax
rates impact no other states.
Table 3 reports the degree distribution across all nodes (states), splitting for in-networks and
out-networks. In , the in-degree is by construction equal to the out-degree, as all ties are
reciprocal. The greater sparsity of the network of economic neighbors is again re‡ected in the
degree distribution being lower for than for . In the dispersion of in- and out-
degree networks is very di¤erent (as measured by the standard deviation), being near double for
the in-degree. This asymmetry in further suggests that some highly focal or in‡uential states
drive tax setting behavior in other states.
Figures 3A and 3B show complete histograms for the in- and out-degree across states. The
histogram on the left is for in-degree, and shows that states under generally have lower
in-degree than under . The states that are in‡uenced by the highest number of other states
are Utah, Pennsylvania and Ohio. The histogram on the right for out-degree, shows the …ve states
described above that do not impact other states (Wisconsin, Vermont, New Mexico, New Jersey
and Minnesota). Delaware is an outlier in‡uential state in its out-degree in determining tax setting
in other states: as discussed below, Delaware is a well-known potential tax haven.32
32Dyreng et al. (2013) …nd that taxes play an important role in determining whether …rms locate subsidiaries in
25
4.6 Factor Mobility or Yardstick Competition?
We conclude by presenting two strategies to shed light on whether factor mobility and yardstick
competition drive these social interactions: (i) exploiting information in the identi…ed social inter-
actions matrix ; (ii) following Besley and Case (1995), using gubernatorial re-election as an
indirect test of the relevance of yardstick competition.
In our …rst strategy we estimate the factors correlated with the existence/strength of links
between states and in using the following dyadic regression:
= 0 + 1 + 2 + 3 +
We discretize link strength so 2 f0 1g and predict the existence of a link using a linear
probability model. We then estimate the correlates of link strength 2 [0 1] using a Tobit
model. The elements and correspond to characteristics of the pair of states ( ), of
state , and state , respectively. Covariates are time-averaged over the sample period, and robust
standard errors are reported. The sample thus corresponds to £ ( ¡ 1) = 48 £ 47 = 2256
potential links that could have formed.
Table 4 presents the results. Column 1 controls only for whether states and are geographic
neighbors. This is highly predictive of a link between them. Columns 2 and 3 show that distance
between states also negatively correlates with them being linked, but that when both geographic
adjacency and distances are included, the former is more predictive. Hence, we control only for
whether and are geographic neighbors in the remaining Columns.
The next set of speci…cations use the insight that economic neighbors are likely to be based
on a mixture of similarity in geography, income per capita, and demography (Case et al., 1989).
Column 4 thus adds two covariates to capture the economic and demographic homophily
between states and . GDP homophily is the absolute di¤erence in the states GDP per capita.
Demographic homophily is the absolute di¤erence of the share of young people (aged 5-17) plus
the absolute di¤erence of the share of elderly people (aged 65+) across the states. GDP homophily
predicts ties, whereas demographic homophily does not.
Columns 5 to 7 then sequentially add in several sets of controls. For labor mobility, we use
net state-to-state migration data to control for the net migration ‡ow of individuals from state
to state (de…ned as the ‡ow from to minus the ‡ow from to ).33 We then add a political
Delaware: a Delaware-based state tax avoidance strategy lowers state e¤ective tax rates by around 1 percentagepoint. They also report that in June 2010, Delaware landed at the top of National Geographic magazine’s publishedlist of the most secretive tax havens in the world (ahead of foreign tax havens such as Luxembourg, Switzerland,and the Cayman Islands).
33We also experimented with alternative measures of labor migration, and results were qualitatively the same.State-to-state migration data are based on year-to-year address changes reported on individual income tax re-turns …led with the IRS. The data cover …ling years 1991 through 2015, and include the number of returns …led,which approximates the number of households that migrated, the number of personal exemptions claimed, whichapproximates the number of individuals who migrated. The data are available at https://www.irs.gov/statistics/soi-
26
homophily variable between states. For any given year, this is set to one if a pair of states have
governors of the same political party. As this is time averaged over our sample, this element
captures the share of the sample period in which the states have governors of the same party.
Lastly, we include whether state is considered a tax haven (and so might have disproportionate
in‡uence on other states). Based on Findley et al. (2012), the following states are coded as tax
havens: Nevada, Delaware, Montana, South Dakota, Wyoming and New York. This corroborates
earlier evidence in Figure 3B, where Delaware, Wyoming and Nevada were among the states with
the highest out-degree.
The speci…cation in Column 7 shows that with this full set of controls, geographic adjacency
remains a robust predictor of the existence of links between states. However, the identi…ed economic
network highlights additional signi…cant predictors of tax competition between states: political
homophily reduces the likelihood of a link, suggesting any yardstick competition driving social
interactions occurs when voters compare their governor to those of the opposing party in other
states. The tax haven states appear to be especially in‡uential in the tax setting behaviors of other
states. The strong in‡uence of tax haven states might lead to a race-to-the-bottom. Relative to
these factors, the economic and demographic similarity between states play an insigni…cant role in
determining interactions between states.
The …nal column considers the continuous link strength as an outcome and reports Tobit partial
average e¤ect estimates. This reinforces that geography, political homophily, and tax haven status
all robustly correlate to the strength of in‡uence states tax setting has on others. Labor mobility
between states does not robustly predict either the existence or strength of ties.
Our second strategy to investigate factor mobility and yardstick competition follows the in-
tuition of Besley and Case (1995). They suggest an indirect test of the relevance of yardstick
competition is that this mechanism only applies to governors not facing term limits. Therefore we
compare our main e¤ects across two subsamples: state-years in which the governor can and cannot
run for reelection. The results are reported in Table 5. The 2SLS results suggest that in both
samples, endogenous social interaction e¤ects exist, although they are more precisely estimated
when governors can run for re-election.
Taken together, our evidence suggests that both factor mobility (of both labor and capital,
as measured through the in‡uence of tax havens), and yardstick competition (occurring through
comparisons to governors of the other party), are important mechanisms driving the existence and
strength of interactions in tax-setting behavior across US states.
Finally, in the Appendix, we contrast how shocks to tax setting in a given state propagate
under the identi…ed interactions matrix , relative to what would have been predicted under a
postulated network structure based on . As is spatially more dispersed than , the
general equilibrium e¤ects might be very di¤erent under the two network structures. We therefore
discuss the implications for tax inequality under and the counterfactual.
tax-stats-migration-data (accessed September 2017).
27
5 Discussion
In a canonical social interactions model, we provide su¢cient conditions under which the social
interactions matrix, and endogenous and exogenous social e¤ects are all globally identi…ed, even
absent information on social links. Our identi…cation strategy is novel, and may bear fruit in
other areas. We describe how high-dimensional estimation techniques can be used to estimate
the model based on the Adaptive Elastic Net GMM method. We showcase our method in Monte
Carlo simulations using two stylized and two real world networks: these highlight that even in
panels as short as = 5, the majority of social ties can be correctly identi…ed. Finally, we
employ this estimation strategy to provide novel insights in a classic social interactions problem:
tax competition across US states.
Our method is immediately applicable to other classic social interactions problems. For ex-
ample, in …nance a long-standing question has been whether CEOs are subject to relative perform-
ance evaluation, and if so, what is the comparison set of …rms/CEOs used (Edmans and Gabaix,
2016).34 Other …elds such as macroeconomics, political economy and trade are all obvious areas
where social interactions across jurisdictions/countries etc. could drive key outcomes, panel data
exist, and the number of nodes is relatively …xed. Our approach can also be applied to new con-
texts where social interactions determine economic behavior but data on social links is absent.
Advances in the availability of administrative data, data from social media or mobile technologies,
and high frequency data in …nance and from online economic transactions, all o¤er new possib-
ilities to identify social interactions. For example, van Vliet, 2018 studies the interconnectedness
between the largest …nancial institutions in during the 2008 …nancial crisis using readlily available
market data, in which = 13 and = 500.
Three further directions for future research are of note. First, under partial observability of
0 (as in Blume et al., 2015), the number of parameters in 0 to be retrieved falls quickly. Our
approach can then still be applied to complete knowledge of 0, and this could be achieved with
potentially weaker assumptions for identi…cation, and in even shorter panels. To illustrate possib-
ilities, Figure 4 shows results from a …nal simulation exercise in which we assume the researcher
starts with partial knowledge of 0. We do so for the Banerjee et al. (2013) village family network,
showing simulation results for scenarios in which the researcher knows the social ties of the three
(…ve, ten) households with the highest out-degree. For comparison we also show the earlier sim-
ulation results when 0 is entirely unknown. This clearly illustrates that with partial knowledge
34Edmans and Gabaix (2016) overview the theory and empirics of executive compensation. Applying the inform-ativeness principle in contract theory to CEO pay suggests peer performance is informative about the degree towhich …rm value is due to high CEO e¤ort or luck. In a …rst generation of studies, Aggarwal and Samwick (1999)and Murphy (1999) showed that CEO pay is determined by absolute, rather than relative performance. However,this conclusion has been challenged by others such as Gong et al. (2011) who argue these conclusions arise fromidentifying relative performance evaluation (RPE) based on an implicit approach, assuming a peer group (e.g. basedon industry and/or size). Indeed, when Gong et al. (2011) study the explicit use of RPE, based on the disclosureof peer …rms and performance measures mandated by the SEC in 2006, they actually …nd that 25% of S&P 1500…rms explicitly using RPE. We are currently working on using our method to provide novel evidence on the matter.
28
of the social network, performance on all metrics improves rapidly for any given .
Second, we have developed our approach in the context of the canonical linear social interactions
model (1). This builds on Manski (1993) when 0 is known to the researcher, and the re‡ection
problem is the main challenge in identifying endogenous and exogenous social e¤ects. However, as
established in Blume et al. (2011) and Blume et al. (2015), the re‡ection problem is functional-form
dependent and may not apply to many non-linear models. An important topic for future research
is thus to extend the insights gathered here to non-linear social interaction settings.
Finally, our approach has taken the network structure as predetermined and …xed. Clearly, an
important part of the social networks literature examines endogenous network formation (Jackson
et al., 2017; de Paula, 2017). Our analysis allows us to begin probing the issue in two ways. First,
the kind of dyadic regression analysis in Section 4 on the correlates of entries in 0 suggests
factors driving link formation and dissolution. Second, it is possible to examine whether the iden-
ti…ed social interactions matrix is stable over time. To illustrate the possibility in a real world
setting, we extend our application on tax competition to investigate the stability of by run-
ning the procedure in two subsamples, each with = 26 periods: 1962-88 and 1989-2015. Panels A
and B in Figure 5 shows the resulting estimated economic networks in each subsample, and Panel
C provides network statistics for each subsample panel (as well as for the earlier estimated eco-
nomic network and the network based on geographic neighbors). This highlights that the network
structure of tax competition has changed over time, with the later sample network from 1989-2015
having fewer edges, fewer reciprocated edges, lower clustering and lower degree distribution.
This analysis leads naturally to a broad agenda going forward, to address the challenge of
simultaneously identifying and estimating time varying models of network formation and social
interaction, all in cases where data on social networks is not required.
References
Acemoglu, D., V. Carvalho, A. Ozdaglar, and A. Tahbaz-Salehi (2012). The Network
Origins of Aggregate Fluctuations. Econometrica, 80, 1977–2016.
Aggarwal, R. K. and A. A. Samwick (1999). Executive Compensation, Strategic Competition,
and Relative Performance Evaluation: theory and evidence. The Journal of Finance, 54.
Ambrosetti, A. and G. Prodi (1972). On the Inversion of Some Di¤erentiable Mappings with
Singularities between Banach Spaces. Annali di Matematica Pura ed Applicata, 93, 231–46.
——— (1995). A Primer of Nonlinear Analysis, Cambridge University Press.
Anselin, L. (2010). Thirty Years of Spatial Econometrics. Papers in Regional Science, 89, 3–25.
29
Atalay, E., A. Hortacsu, J. Roberts, and C. Syverson (2011). Network Structure of
Production. Proceedings of the American Mathematical Society, 108, 5199–202.
Ballester, C., A. Calvo-Armendol, and Y. Zenou (2006). Who’s Who in Networks.
Wanted: The Key Player. Econometrica, 74, 1403–17.
Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. O. Jackson (2013). The Di¤usion
of Micro…nance. Science, 341, 1236498.
Besley, T. and A. Case (1994). Unnatural Experiments? Estimating the Incidence of Endo-
genous Policies. NBER Working Paper 4956.
——— (1995). Incumbent Behavior: Vote-seeking, Tax-setting, and Yardstick Competition. Amer-
ican Economic Review, 85, 25–45.
Blume, L., W. A. Brock, S. N. Durlauf, and Y. Ioannides (2011). Identi…cation of Social
Interactions. in Handbook of Social Economics, ed. by J. Behabib, A. Bisin, and M. O. Jackson,
Noth-Holland, vol. 1B.
Blume, L. E., W. A. Brock, S. N. Durlauf, and R. Jayaraman (2015). Linear Social
Interactions Models. Journal of Political Economy, 123, 444–96.
Bonaldi, P., A. Hortacsu, and J. Kastl (2015). An Empirical Analysis of Funding Costs
Spillovers in the EURO-zone with Application to Systemic Risk. Princeton University Working
Paper.
Bramoullé, Y., H. Djebbari, and B. Fortin (2009). Identi…cation of Peer E¤ects Through
Social Networks. Journal of Econometrics, 150, 41–55.
Brennan, G. and J. Buchanan (1980). The Power to Tax: Analytical Foundations of a Fiscal
Constitution, Cambridge University Press.
Breza, E., A. Chandrasekhar, T. McCormick, and M. Pan (2017). Using Aggregated Re-
lational Data to Feasibly Identify Network Structure without Network Data. Harvard University
Working Paper.
Bursztyn, L., F. Ederer, B. Ferman, and N. Yuchtman (2014). Understanding Mech-
anisms Underlying Peer E¤ects: Evidence From a Field Experiment on Financial Decisions.
Econometrica, 82, 1273–301.
Caner, M. and H. H. Zhang (2014). Adaptive Elastic Net for Generalized Method of Moments.
Journal of Business and Economic Statistics, 32, 30–47.
30
Case, A., J. R. Hines, and H. S. Rosen (1989). Copycatting: Fiscal Policies of States and
Their Neighborrs. NBER Working Paper 3032.
Case, A., H. Rosen, and J. Hines (1993). Budget Spillovers and Fiscal Policy Interdependence:
Evidence from the States. Journal of Public Economics, 52, 285–307.
Chandrasekhar, A. and R. Lewis (2016). Econometrics of Sampled Networks. Working Paper.
Chaney, T. (2014). The Network Structure of International Trade. American Economic Review,
104, 3600–34.
Chiappori, P.-A., I. Komunjer, and D. Kristensen (2015). Nonparametric Identi…cation
and Estimation of Transformation Models. Journal of Econometrics, 188, 22–39.
Coleman, J. S. (1964). Introduction to Mathematical Sociology, London Free Press Glencoe.
Conley, T. G. and C. R. Udry (2010). Learning About a New Technology: Pineapple in
Ghana. American Economic Review, 100, 35–69.
De Giorgi, G., M. Pellizzari, and S. Redaelli (2010). Identi…cation of Social Interactions
through Partially Overlapping Peer Groups. American Economic Journal: Applied Economics,
2, 241–75.
de Marco, G., G. Gorni, and G. Zampieri (2014). Global Inversion of Functions: an Intro-
duction. ArXiv:1410.7902v1.
de Paula, A. (2017). Econometrics of Network Models. in Advances in Economics and Econo-
metrics: Theory and Applications, ed. by B. Honore, A. Pakes, M. Piazzesi, and L. Samuelson,
Cambridge University Press.
Devereux, M. and R. Griffith (1998). Taxes and the Location of Production: Evidence from
a Panel of US Multinationals. Journal of Public Economics, 68, 335–67.
Diebold, F. X. and K. Yilmaz (2015). Financial and Macroeconomic Connectedness: A Net-
work Approach to Measurement and Monitoring, Oxford University Press.
Dyreng, S., B. Lindsey, and J. Thornbock (2013). Exploring the Role Delaware Plays as a
Domestic Tax Haven. Journal of Financial Economics, 108, 751–72.
Edmans, A. and X. Gabaix (2016). Executive Compensation: a modern primer. Journal of
Economic literature, 54, 1232–87.
Erdos, P. and A. Renyi (1960). On the Evolution of Random Graphs. Publ. Math. Inst. Hung.
Acad. Sci, 5, 17–60.
31
Farber, S., A. Páez, and E. Volz (2009). Topology and dependency tests in spatial and
network autoregressive models. Geographical Analysis, 41, 158–180.
Findley, M., D. Nielson, and J. Sharman (2012). Global Shell Games: Testing Money
Launderers’ and Terrorist Financiers’ Access to Shell Companies. Gri¢th University WP.
Gautier, E. and C. Rose (2016). Inference in Social E¤ects when the Network is Sparse and
Unknown. (in preparation).
Gautier, E. and A. Tsybakov (2014). High-Dimensional Instrumental Variables Regression
and Con…dence Sets. Working Paper CREST.
Gong, G., L. Y. Li, and J. Y. Shin (2011). Relative Performance Evaluation and Related Peer
Groups in Executive Compensation Contracts. The Accounting Review, 86, 1007–43.
Granovetter, M. (1973). The Strength of Weak Ties. American Journal of Sociology, 6, 1360–
80.
Hines, J. (1996). Altered States: Taxes and the Location of Foreign Direct Investment in America.
American Economic Review, 86, 1076–94.
Horn, R. A. and C. R. Johnson (2013). Matrix Analysis, Cambridge University Press.
Jackson, M., B. Rogers, and Y. Zenou (2017). The Economic Consequences of Social Net-
work Structure. Journal of Economic Literature, 55, 49–95.
Janeba, E. and S. Osterleh (2013). Tax and the City – A Theory of Local Tax Competition.
Journal of Public Economics, 106, 89–100.
Kennedy, J. and R. Eberhart (1995). Particle Swarm Optimization. Proceedings of the IEEE
International Conference on Neural Networks, IV, 1942–8.
Kleven, H., C. Landais, and E. Saez (2013). Taxation and International Mobility of Super-
stars: Evidence from the European Football Market. American Economic Review, 103, 1892–924.
——— (2014). Migration and Wage E¤ects of Taxing Top Earners: Evidence from the Foreigners’
Tax Scheme in Denmark. Quarterly Journal of Economics, 129, 333–78.
Kline, B. and E. Tamer (2016). Bayesian Inference in a Class of Partially Identi…ed Models.
Quantitative Economics, 7, 329–366.
Komunjer, I. (2012). Global Identi…cation in Nonlinear Models with Moment Restrictions. Eco-
nometric Theory, 28, 719–29.
Krantz, S. G. and H. R. Parks (2013). The Implicit Function Theorem, Birkhauser.
32
Lam, C. and P. C. Souza (2016). Detection and Estimation of Block Structure in Spatial Weight
Matrix. Econometric Reviews, 35, 1347–1376.
——— (2019). Estimation and Selection of Spatial Weight Matrix in a Spatial Lag Model. Journal
of Business and Economic Statistics, forthcoming.
Lee, L.-F. (2004). Asumptotic Distributions of Quasi-Maximum Likelihood Estimators for Spatial
Autoregressive Models. Econometrica, 72, 25.
——— (2007). Identi…cation and Estimation of Econometric Models with Group Interactions,
Contextual Factors and Fixed E¤ects. Journal of Econometrics, 60, 531–42.
Lee, S. and A. Lewbel (2013). Nonparametric Identi…cation of Accelerated Failure Time Com-
peting Risks Models. Econometric Theory, 29, 905–19.
Lewbel, A., X. Qu, and X. Tan (2019). Social Networks with Misclassi…ed or Unobserved
Links. Manuscript.
Manresa, E. (2016). Estimating the Structure of Social Interactions Using Panel Data. Manu-
script.
Manski, C. F. (1993). Identi…cation of Endogenous Social E¤ects: the re‡ection problem. The
Review of Economic Studies, 60, 531–42.
Meinshausen, N. and P. Buhlmann (2006). High-Dimensional Graphs and Variable Selection
with the Lasso. The Annals of Statistics, 34, 1436–1462.
Mizruchi, M. S. and E. J. Neuman (2008). The e¤ect of density on the level of bias in the
network autocorrelation model. Social Networks, 30, 190–200.
Murphy, K. J. (1999). Executive Compensation. Handbook of Labor Economics, 3, 2485–563.
Neuman, E. J. and M. S. Mizruchi (2010). Structure and bias in the network autocorrelation
model. Social Networks, 32, 290–300.
Oates, W. and R. Schwab (1988). Economic Competition Among Jurisdictions: E¢ciency-
enhancing or Distortion-inducing?. Journal of Public Economics, 35, 333–54.
Rose, C. (2015). Essays in Applied Microeconometrics. Ph.D. thesis, University of Bristol.
Rothenberg, T. (1971). Identi…cation in Parametric Models. Econometrica, 39, 577–91.
Rothenhäusler, D., C. Heinze, J. Peters, and N. Meinshausen (2015). BACKSHIFT:
Learning causal cyclic graphs from unknown shift interventions. in Advances in Neural Inform-
ation Processing Systems, 1513–1521.
33
Sacerdote, B. (2001). Peer E¤ects with Random Assingment: Results for Dartmouth Room-
mates. The Quarterly Journal of Economics, 116, 681–704.
Shleifer, A. (1985). A Theory of Yardstick Competition. Rand Journal of Economics, 16.
Smith, T. E. (2009). Estimation bias in spatial models with strongly connected weight matrices.
Geographical Analysis, 41, 307–332.
Souza, P. C. (2014). Estimating Network E¤ects without Network Data. PUC-Rio Working
Paper.
Tibshirani, R. J. and J. Taylor (2012). Degrees of freedom in lasso problems. The Annals of
Statistics, 40, 1198–1232.
Tiebout, C. (1956). A Pure Theory of Local Expenditures. Journal of Political Economy, 64,
416–24.
van Vliet, W. (2018). Connections as Jumps: Estimating Financial Interconnectedness from
Market Data. Manuscript.
Wang, W., E. J. Neuman, and D. A. Newman (2014). Statistical power of the social network
autocorrelation model. Social Networks, 38, 88–99.
Watts, D. J. (1999). Networks, Dynamics, and the Small-World Phenomenon. American Journal
of Sociology, 105, 493–527.
Wilson, J. (1986). A Theory of Interregional Tax Competition. Journal of Urban Economics, 19,
296–315.
——— (1999). Theories of Tax Competition. National Tax Journal, 52, 269–304.
Yuan, M. and Y. Lin (2007). Model Selection and Estimation in the Gaussian Graphical Model.
Biometrika, 94, 19–35.
Zou, H. (2006). The Adaptive Lasso and Its Oracle Properties. Journal of the American Statistical
Association, 101, 1418–29.
Zou, H. and T. Hastie (2005). Regularization and Variable Selection via the Elastic Net. Journal
of the Royal Statistical Society: Series B (Statistical Methodology), 67, 301–320.
Zou, H., T. Hastie, and R. Tibshirani (2007). On the “Degrees of Freedom” of the LASSO.
The Annals of Statistics, 35, 2173–2192.
Zou, H. and H. H. Zhang (2009). On the Adaptive Elastic-net with a Diverging Number of
Parameters. Ann. Statist., 37, 1733–51.
34
A Proofs
Theorem 1
Proof. The local identi…cation result follows Rothenberg (1971). Under the assumptions in our
model, the parameter space £ ½ R is an open set (recall that = ( ¡ 1) + 3.) This
corresponds to Assumption I in Rothenberg (1971).
We have that,
¦
= ( ¡ )¡1¢ ( ¡ )¡1 ( + ) + ( ¡ )¡1 ¢
¦
= ( ¡ )¡1 ( ¡ )¡1 ( + )
¦
= ( ¡ )¡1
¦
= ( ¡ )¡1
where ¢ is the £ matrix with 1 in the ( )-th position and zero elsewhere. Write the
2 £ derivative matrix r¦ ´ vec(¦)0
. By assumption, row in matrix sums up to one,
incorporated through the restriction that ´P
=1 6= ¡ 1 = 0, for the unit-normalised row
. The derivative of the restriction is the -dimensional vector r0 ´
0=
£0 0¡1 01£3
¤
(where is an -dimensional vector with 1 in the th component and zero, otherwise). Following
Theorem 6 of Rothenberg (1971), the structural parameters 2 £ are locally identi…ed if, and
only if, the matrix r ´ [r0¦ r0
]0has rank .35
If r is does not have rank , there is a nonzero vector c ´¡12 ¡1
¢0
such
that r ¢ c = 0. This implies that
12
¦
12
+ ¢ ¢ ¢+ ¡1
¦
¡1
+ ¦
+
¦
+
¦
= 0 (12)
and, for the unit-normalized row (see A4),
X
6==1
= 0 (13)
35For a parameter vector to be locally identi…ed, Rothenberg (1971) requires that the derivative matrix r haverank at that point and that this vector be (rank-)regular. A (rank-)regular point of the parameter space is onefor which there is a neighborhood where the rank of r is constant (see De…nition 4 in Rothenberg, 1971). Becausewe show that the derivative matrix has rank at every point in the parameter space, this also guarantees thatevery point in the parameter space is (rank-)regular.
35
Premultiplying equation (12) by ( ¡ ) and substituting the derivatives,
X
=16=
£¢ ( ¡ )¡1 ( + ) + ¢
¤+
+ ( ¡ )¡1 ( + ) + + = 0
De…ne ´P
=16= ¢. Since the spectral radius of is strictly less than one by A2, one
can show (by representing ( ¡ )¡1 as a Neumann series, for instance) that ( + ) and
( ¡ )¡1 commute. Then, the expression above is equivalent to
( + ) ( ¡ )¡1 + + ( + ) ( ¡ )¡1 + + = 0
Post-multiplying by ( ¡ ), we obtain
( + ) + ( ¡ ) + ( + ) + ( ¡ ) + ( ¡ ) = 0
which, upon rearrangement, yields
( + ) + + ( ¡ + ) + ( ¡ )2 = 0 (14)
Because = 0 and = 0 (by A1), we have that +( ¡ ) (2) = 0 for all = 1 .
Since by assumption A5 there isn’t a constant such that diag ( 20 ) = , then = ¡ = 0.
Plugging back in (14), we obtain
( + ) + ( + ) = 0
which implies that = ¡++
since + 6= 0 by assumption A3. Taking the sum of the
elements in row , we get
( + )X
6==1
+ ( + ) = 0
Note that, by equation (13),P
6==1 = 0. So + = 0 and = ¡++
= 0. This
implies that = 0 for any and . Combining + = 0 with ¡ = 0 obtained above,
we get that ( + ) = 0. Since + 6= 0, then = 0. Given that + = 0, it follows
that = 0. This shows that 2 £ is locally identi…ed.
Corollary 1
Proof. The parameter 0 being locally identi…ed (see Theorem 1) implies that the set f : ¦() =
¦(0)g is discrete. If it is also compact, then the set is …nite. To establish that we now show that
36
¦ is a proper function: the inverse image ¦¡1() of any compact set ½ R is also compact
(see Krantz and Parks, 2013, p.124).
Let A be a compact set in the space of £ real matrices. Since it is a compact set in
a …nite dimensional space, it is closed and bounded. Since ¦ is a continuous function of , the
pre-image of a compact set, which is closed, is also closed. Because W is bounded and 2 (¡1 1),
their corresponding coordinates in 2 ¦¡1(A) are bounded. Suppose the coordinates for or in
2 ¦¡1(A) are not bounded. So one can …nd a sequence ()1=1 such that jj ! 1 or jj ! 1.
Denote the Frobenius norm of the matrix as kk. By the submultiplicative property kk ·
kk ¢ kk,
k + k ·°°( + ) ( ¡ )¡1
°° ¢ k ¡ k
Note that ( ¡ )¡1 and ( + ) commute, and so
°°( + ) ( ¡ )¡1
°° =
°°( ¡ )¡1 ( + )
°° = k¦k
It follows that
k + k
k ¡ k· k¦k
Given has zero main diagonal,
k + k2 = 2 kk2 + 2 kk2 = 2 + 2 kk2
Also, k ¡ k2 = + 2 kk2 · + 2, for some constant 2 R, since W is bounded by
assumption A2. We then have that
q2 + 2 kk2
p + 2
· k¦k
Since jj 1 by Assumption (A2) the denominator above is bounded. Hence jj ! 1 )
k¦()k ! 1. We now use the fact thatP
= 1 to show that there is a lower bound on
kk2, and so jj ! 1 ) k¦()k ! 1. To see this, note that
mins.t.
=1
kk2 ¸ mins.t.
=1
X
=1
2
37
The Lagrangean for the right-hand side minimization problem is:
L (1 ¡1+1 ;) =X
=1
2 ¡
ÃX
¡ 1
!
where is the Lagrangean multiplier for the normalisation constraint. The …rst-order conditions
for this convex minimization problem are:
L
= 2 ¡ = 0 for any 6=
L
=
X
=1
¡ 1 = 0
The …rst equation implies that =2
for = 1 ¡ 1 + 1 . Using the fact that
= 0, the second equation implies that = 2( ¡ 1). We have then that =1
¡1 6=
and, consequently, kk2 ¸ ( ¡ 1) 1(¡1)2
= 1¡1
. Hence, if jj ! 1, the numerator in the lower
bound for k¦k above also goes to in…nity. Consequently, A would not be compact.
Therefore, if A is compact the coordinates in 2 ¦¡1(A) corresponding to and are also
bounded. Hence, ¦¡1(A) is bounded (and closed). Consequently it is compact.
For a given reduced form parameter matrix ¦, the set f : ¦() = ¦(0)g is then compact.
Since it is also discrete, it is …nite.
The following lemmas are used in proving Theorem 2.
Lemma 1. Assume (A1)-(A5). If 0 = 0, 0 is such that (0)12 = (0)21 = 1 and (0) = 0
otherwise, with 0 6= 0 and 0 6= 0, then 0 2 £ is identi…ed.
Proof. Take = (12 ¡1 ) 2 £ possibily di¤erent from 0 such that the models
are observationally equivalent, so ¦0 = ¦. Then
( ¡ 00)¡1(0 + 00) = ( ¡ )¡1( + )
Since 0 = 0 and ( ¡ )¡1 and ( + ) commute (see the proof for Theorem 1), it follows
that
¦0 = ¦ , 0( ¡ 00)¡1 = ( + )( ¡ )¡1
or, equivalently,
0( ¡ ) = ( ¡ 00)( + )
38
This last equation implies that
(0 ¡ ) ¡ ( + 0) + 00 + 00 = 0 (15)
We …rst note that (0 ) = 0 since (0) = 0 for any 1 · · and, by Assumption (A1),
( ) = (0) = 0. So 0 = . Taking elements ( ) such that ¸ 3 and 6= in equation
(15), and using the fact that 0 = , we …nd that ¡( + 0)( ) = ¡( + )( ) = 0 for any
( ) such that ¸ 3 and 6= . By Assumption (A3), + 6= 0 and it follows that ( ) = 0
for any ( ) such that ¸ 3 and 6= . In fact, since ( ) = 0 by Assumption (A1), we get that
( ) = 0 for any ( ) such that ¸ 3.
Using Assumption (A1) and since 0 = , elements (1 1) and (2 2) in equation (15) imply
that 0( )21 = 0( )12 = 0. Given that 0 6= 0, we get that ( )21 = ( )12 = 0.
From element (1 2) in equation (15) we …nd that ¡( + 0)( )12 + 0 = 0 or, equivalently,
(0¡( )12)0¡( )12 = 0. Given that ( )12 = 0 and 0 6= 0, it must be that 0¡( )12 =
0. Making the analogous argument for element (2 1), we would also obtain that 0 ¡ ( )21 = 0.
If both ( )12 and ( )21 are equal to zero, using the fact that = 0 for any ( ) such that
¸ 3, we would then obtain that 2 is equal to
2
66666664
0 0 ( )13 ¢ ¢ ¢ ( )1
0 0 ( )23 ¢ ¢ ¢ ( )2
0 0 0 ¢ ¢ ¢ 0...
......
. . ....
0 0 0 ¢ ¢ ¢ 0
3
77777775
2
=
2
66666664
0 0 0 ¢ ¢ ¢ 0
0 0 0 ¢ ¢ ¢ 0
0 0 0 ¢ ¢ ¢ 0...
......
. . ....
0 0 0 ¢ ¢ ¢ 0
3
77777775
which contradicts Assumption (A5). Hence ( )12 6= 0 or ( )21 6= 0. If ( )12 6= 0, using the
fact that ( )12 = 0, we get that = 0. Equivalently, if ( )21 6= 0, and using the fact that
( )21 = 0, we again get that = 0. So, in either case, = 0 = 0.
Taking element (1 ) in equation (15), with ¸ 3, we get that ¡( + 0)1 + 0( )2 =
¡01 = 0. Similarly, element (2 ), with ¸ 3 implies that ¡( + 0)2 + 0( )1 =
¡02 = 0. Then, from ¡0( )1 = ¡0( )2 = 0 for ¸ 3, it follows that ¡( )1 =
¡( )2 = 0 since 0 6= 0.
From 0 ¡ ( )12 = 0, if ( )12 6= 0, we get that = 0( )12 6= 0. Equivalently, if
( )21 6= 0, we get that = 0( )21 6= 0. Since ( )12 6= 0 or ( )21 6= 0, we obtain that 6= 0.
Then, because ¡( )1 = ( )2 = 0 for ¸ 3, we have that ( )1 = ( )2 = 0 for ¸ 3.
Given that 0¡( )12 = 0, 0¡( )21 = 0 and 6= 0, we obtain that ( )12 = ( )21 =0.
Since ( )1 = 0 for 6= 2, ( )2 = 0 for 6= 1 and ( ) = 0 for ¸ 3, by Assumption
(A5) we get that ( )12 = ( )21 = 1 and = 0. Hence, (( )12 ( )¡1 ) =
((0)12 (0)¡1 0 0 0).
39
Lemma 2. Assume (A1)-(A2) and (A4)-(A5). The image of ¦(¢), for 2 £+, is path-connected
and, therefore, connected.
Proof. Take and ¤ 2 £+. Consider …rst the subvectors corresponding to the adjacency matrices
and ¤. Without loss of generality, let 1 be ordered such that ( 2)11 (2)22. Con-
sider the adjacency matrix ¤ corresponding to the network of directed connections f(1 2) (2 1)g
and f(3 4) (4 5) ( ¡ 1 ) ( 3)g:
¤ =
2
66666664
0 1 0 0 ¢ ¢ ¢ 0
1 0 0 0 ¢ ¢ ¢ 0
0 0 0 1 ¢ ¢ ¢ 0...
......
. . . . . ....
0 0 1 0 ¢ ¢ ¢ 0
3
77777775
Note that diag( 2¤ ) = (1 1 0 0) and this is an admissible adjacency matrix under assumptions
(A1)-(A2) and (A4)-(A5). We …rst show that is path-connected to ¤.
Consider the path given by
() = ¤ + (1¡ )
which implies that
( ()2)11 = (1¡ )2( 2)11 + 2 + (1¡ )(12 +21)
( ()2)22 = (1¡ )2( 2)22 + 2 + (1¡ )(12 +21)
Since ( ()2)11 ¡ ( ()2)22 = (1 ¡ )2[( 2)11 ¡ ( 2)22] 0 for 2 [0 1) and (1) = ¤,
(A5) is satis…ed for any matrix () such that 2 [0 1]. Since all rows in ¤ sum to one and
(¤) = 0 for any , it is straightforward to see that () also satis…es (A1) and (A4). Finally,P
=1 j()j · P
=1 j(¤)j+ (1¡ )P
=1 jj · 1 for every = 1 and () satis…es
Assumption (A2).
If ¤ is such that ( ¤2)11 6= ( ¤2)22, the convex combination of ¤ and ¤ is also seen to
satisfy (A1)-(A2) and (A4)-(A5) and a path between and ¤ can be constructed via ¤. If,
on the other hand ( ¤2)11 = (¤2)22, suppose without loss of generality that ( ¤2)11 6= ( ¤2)33.
In this case, one can construct a path between ¤ and ¤¤ where ¤¤ represents the network of
40
directed connections f(1 3) (3 1)g and f(2 4) (4 5) ( ¡ 1 ) ( 2)g:
¤¤ =
2
66666664
0 0 1 0 ¢ ¢ ¢ 0
0 0 0 1 ¢ ¢ ¢ 0
1 0 0 0 ¢ ¢ ¢ 0...
......
. . . . . ....
0 1 0 0 ¢ ¢ ¢ 0
3
77777775
Like () above, this path can be seen to satisfy assumptions (A1)-(A2) and (A4)-(A5). Now
note that a path can also be constructed between ¤ and ¤¤ as their convex combination also
satis…es (A1)-(A2) and (A4)-(A5). For example, note that () = ¤ + (1¡ )¤¤ is such that
( ()2)11 = 2 + (1 ¡ )2 and ( ()2) = 0 so ( ()2)11 ¡ ( ()2) 0 for any 2 (0 1)
and both (0) and (1) satisfy (A5). Hence, we can construct a path () between and ¤
through ¤ and ¤¤.
Furthermore, () = ¤+(1¡), () = (¤¤+(1¡))(¤+(1¡)), () = ¤+(1¡)
are such that
() ´ ()() + () = (¤¤ + ¤) + (1¡ )( + ) 0
since ¤ and 2 £+. (Note also that j()j 1 so Assumption (A2) is satis…ed.) These facts taken
together imply that
() ´ ( ()12 ()¡1 () () ()) 2 £+
That is, £+ is path-connected and therefore connected. Since ¦(¢) is continuous on £+, ¦(£+) is
connected.
Theorem 2
Proof. The proof uses Corollary 1.4 in Ambrosetti and Prodi (1995, p. 46),36 which we reproduce
here with our notation for convenience: Suppose the function ¦(¢) is continuous, proper and locally
invertible with a connected image. Then the cardinality of ¦¡1(¦) is constant for any ¦ in the
image of ¦(¢).
The mapping ¦() is continuous and proper (by Corollary 1), with connected image (Lemma
2), and non-singular Jacobian at any point (as per the proof for Theorem 1) which guarantees local
invertibility. Following Corollary 1.4 in Ambrosetti and Prodi (1995, p.46) reproduced above, we
obtain that the cardinality of the pre-image of ¦() is …nite and constant. Take 2 £+ such that
36Related results can be found in Ambrosetti and Prodi (1972) and de Marco et al. (2014)
41
= 0, ( )12 = ( )21 = 1 and ( ) = 0 otherwise, with 6= 0 and 6= 0. By Lemma 1, that
cardinality is one.
Corollary 3
Proof. Since 2 (0 1) and ¸ 0,P1
=1 ¡1 is a non-negative matrix. By (5), the o¤-
diagonal elements of ¦() are equal to the o¤-diagonal elements of ( + )P1
=1 ¡1 , the
sign of those elements identi…es the sign of + . By Theorem 2, the model is identi…ed.
Corollary 4
Proof. Since 0 is non-negative and irreducible, there is a real eigenvalue equal to the spectral
radius of 0 corresponding to the unique eigenvector whose entries can be chosen to be strictly
positive (i.e., all the entries share the same sign). A generic eigenvalue of 0, 0, corresponds to
an eigenvalue of ¦0 according to:
¦0 = 0 + (00 + 0)0
1¡ 00
If 0 = 0 + 0 where 0 0 2 R and =p
¡1, then
¦0 = 0 + (00 + 0)0(1¡ 00)¡ 0
20
(1¡ 00)2 + 20
20
+ (00 + 0)0
(1¡ 00)2 + 20
20
If the eigenvalue 0 is real, 0 = 0 and the corresponding ¦0 eigenvalue is also real. Di¤erentiating
(¦0), the real part of ¦0 , with respect to (0) = 0, we get:
(¦0)
0=
(1¡ 00)2 ¡ 20
20
[(1¡ 00)2 + 20
20]2
£ (00 + 0) (16)
If the eigenvalue 0 is real, the expression (16) becomes:
(¦0)
0=
¦00
=1
(1¡ 00)2
£ (00 + 0)
The fraction multiplying 00 + 0 is positive. If 00 + 0 0, the real eigenvalues of ¦0 are
decreasing on the real eigenvalues of 0. Consequently, the eigenvector corresponding to the
largest (real) eigenvalue of 0 will be associated with smallest real eigenvalue of ¦0. If, on the
other hand, 00 + 0 0 the eigenvector corresponding to the largest real eigenvalue of 0 will
correspond to the largest real eigenvalue of ¦0. Since that eigenvector is the unique eigenvector
that can be chosen to have strictly positive entries, the sign of 00 + 0 is identi…ed by the
¦0 eigenvalue it is associated with and whether it is the largest or smallest real eigenvalue. By
Theorem 2, the model is identi…ed.
42
If there is only one real eigenvalue, note that the denominator in the fraction in (16) is positive.
The minimum value of the numerator subject to j0j2 = 20 + 20 · 1 is given by
min00
(1¡ 00)2 ¡ 220 s.t. 20 + 20 · 1
The Lagrangean for this minimization problem is given by:
L (0 0;) = (1¡ 00)2 ¡ 220 + (20 + 20 ¡ 1)
where is the Lagrange multiplier associated with the constraint 20 + 20 · 1. The Kuhn-Tucker
necessary conditions for the solution (¤0 ¤0
¤) of this problem are given by:
(0 :) 0(1¡ 0¤0)¡ ¤¤0 = 0
(0 :) (20 ¡ ¤)¤0 = 0
¤(¤20 + ¤20 ¡ 1) = 0
¤20 + ¤20 · 1 and ¤ ¸ 0
Let 0 6= 0. (Otherwise, the objective function above is equal to one irrespective of 0 or 0 and the
partial derivative is 00 + 0.) If ¤ = 0, 0 implies that ¤0 = 0. Then 0 would have ¤0 = ¡10which violates ¤20 + ¤20 · 1.
Hence, a solution should have ¤ 0. In this case, there are two possibilities: ¤0 = 0 or
¤0 6= 0. If ¤0 6= 0, condition 0 implies that ¤ = 20 and 0 then gives ¤0 = (20)¡1. Because the
constraint is binding, ¤20 = 1¡ (420)¡1. In this case, ¤20 · 1 and ¤20 ¸ 0 requires that j0j ¸ 12.
The value of the minimised objective function in this case 12¡20. This is positive if j0j p22.
The other possibility is to have 0 = 0. Because the constraint is binding, 0 = 1 and the
objective function takes the value (1¡ 0)2 0. Since (1¡ 0)
2 ¡ 12+ 20 = 220 ¡ 20+ 12 ¸ 0,
this solution is dominated by the previous one when j0j ¸ 12.
Consequently, the fraction multiplying 00+0 is non-negative and it can be ascertained that
sgn
·(¦0)
0
¸
= sgn[00 + 0]
as long as j0j p22.
If 00 + 0 0, the real part of the eigenvalues of ¦0 is decreasing on the real part of the
eigenvalues of 0. Consequently, the eigenvector corresponding to the eigenvalue of 0 with the
largest real part will correspond to the eigenvalue of ¦0 with the smallest real part. If, on the
other hand, 00 + 0 0 the eigenvector corresponding to the eigenvalue of 0 with the largest
real part will correspond to the eigenvalue of ¦0 with the largest real part. Since that eigenvector
is the unique eigenvector that can be chosen to have strictly positive entries, the sign of 00+ 0
43
is identi…ed by the ¦0 eigenvalue it is associated with.
By Theorem 2, the model is identi…ed.
Proposition 1
Proof. From equation (6) we observed that ¦0 = ¦0, where is an eigenvector of both 0
and ¦0 with corresponding eigenvalue ¦0 =0+001¡00
. De…ning as the row-sum of ¦0, we also
have that
~¦0( ¡) = ( ¡ )¦0( ¡) = ( ¡ )¦0 ¡ ( ¡ )¦0
= ¦0( ¡) ¡ ( ¡ ) = ¦0( ¡) ¡ ( ¡2)
= ¦0( ¡) ¡ ( ¡) = ¦0( ¡)
where the third equality obtains from¦0 = and the …fth equality holds since is idempotent.
So ~¦0 and ¦0 have common eigenvalues, with corresponding eigenvector ~ = ¡¹ for ~¦0, where
¹ =10, = 1 . Since ¦0 and ~ are observed from ~¦0, identi…cation of ¦0 is equivalent
to identi…cation of ¹ (given diagonilizability).
To establish identi…cation of ¹, note that 0(~+¹) = 0(~+¹) since is an eigenvector
of 0. Consider an alternative constant ¹¤ 6= ¹ that satis…es the previous equation. Then
0(¹ ¡ ¹¤ ) = 0(¹ ¡ ¹¤ )
Since 0 = , must satisfy (1¡ 0)(¹ ¡ ¹¤ ) = 0. For = 2 , j0j 1. So ¹ = ¹¤ and
therefore identi…ed. For = 1, it is known that 1 = 1 with eigenvector 1 = .
Proposition 2
Proof. Under row-sum normalization and j0j 1, ( ¡ 00)¡1 = + 00 + 20
20 + ¢ ¢ ¢ =
+0+20+ ¢ ¢ ¢ = 11¡0
, so ¦01 ´ ( ¡0 )¡1 has constant row-sums. If row-sum normalization
fails, ¦01 may not have constant row-sums. De…ne as the ()-th element of ~. The …rst row
of the system ( ¡ ~)( ¡ 0 )¡1 = ( ¡ ~)0 = 0 is ¤1101 ¡ 1202 ¡ ¢ ¢ ¢ ¡ 10 = 0
where ¤11 = 1¡11 and 0 is the -th element of 0 . If there are possible 0, (1)0
()0 ,
such that [
(1)0
¢ ¢ ¢
()0] has rank , then ¤11 = 12 = ¢ ¢ ¢ = 1 = 0. Since the same reasoning
applies to all rows, ~ is the trivial transformation ~ = .
44
B Estimation
B.1 Sparsity of 0 and ¦0
De…ne ~ as the number of non-zero elements of ¦0. We say that ¦0 is sparse if ~ ¿ . Denote
the number of connected pairs in 0 via paths of any length as ~. We equivalently say that 0
is "sparse connected" if ~ ¿ . We show that sparsity of ¦0 is related to sparse connectedness
of 0.
Proposition 3. ¦0 is sparse if, and only if, the number of unconnected pairs 0 is large.
Proof. For j0j 1, we have that
¦0 = 0 + (00 + 0)1X
=1
¡10 0
Given that 00 + 0 6= 0, it follows directly that [¦0] = 0 if, and only if, there are no paths
between and in0. Therefore, sparsity of ¦0 translates into a large number of ( ) unconnected
pairs in 0.
On the one hand, sparsity does not imply sparse connectedness. A circular graph is clearly
sparse, but all nodes connect with all other nodes through a path of length at most 2. On the
other hand, the sparse connectedness implies sparsity and therefore is a stronger requirement.
To see this, take any arbitrary network with ~ () non-zero elements and ~ () connected
pairs. Now consider the operation of “completing” : for every connected ( ) pair, add a direct
link between ( ) if non-existent in and denote the resulting matrix as C (). It is clear that
~ () · ~ (C ()). Yet, ~ (C ()) = ~ ().
B.2 Adaptive Elastic Net
B.2.1 Implementation and Initial Conditions
To make our procedure robust to the choice of initial condition, we use the particle swarm al-
gorithm. This is an optimization algorithm tailored to more aptly …nd global optima, which does
not depend on choice of initial conditions. It works as follows. The procedure starts from a large
number of initial conditions covering the parameter space, known as “particles” (Kennedy and
Eberhart, 1995). Each particle is iterated independently until convergence. The algorithm returns
45
the optimum calculated across particles.37 38
To ensure compliance with row-sum normalization for each row of , one non-zero parameter
¤ is set to 1 ¡P
=1 6=¤ . This avoids making use of constrained optimization routines.39
We also impose the restriction that ¸ 0 and ¸ 0 by minimizing the objective function with
respect to ~ with = ~2 and ~ with = ~ 2.
Optimization of (8) starts from the initial condition selected by the particle swarm algorithm
and is minimized with respect to the parameters that were neither set to zero nor were chosen to
ensure row-sum normalization. Estimates from the …rst stage are subsequently used to adjust the
penalization, as in the Adaptive Elastic Net GMM objective function (10).40 The steps above are
repeated for di¤erent combinations of = (1 ¤1 2), selected on a grid. The …nal estimate is the
one that minimizes the BIC criterion.
B.3 OLS
For the purpose of estimation, it is convenient to write the model in the stacked form. Let
= [1 ]0 be the £ matrix of explanatory variables, = [1 ]
0 be the £ 1
vector of response variables for individual and 0 = [01
0 ]
0where 0 is a short notation
for the ( )-th element of ¦0. The concise model is then,
= 0 + (17)
for each = 1 , where also = [1 ]0. Model (17) can then be estimated equation-
by-equation. Denote 0 = [010 0
0]0. Stacking the full set of equations,
= 0 + (18)
37We set Caner and Zhang’s (2014) suggestion for the initial condition as one of those particles, with minormodi…cations. The authors suggest calculating the absolute value of the derivative of the unpenalized GMMobjective function evaluated at zero, r , and the set parameters smaller than 1 at zero. The rationale is thatif the GMM objective function is invariant with respect to certain parameters, the Elastic Net problem achieves acorner solution (where parameters are set to zero). In our case, allowing only for positive interactions, we set to zerothe elements such that ¡r · 1. All other elements of gain equal weights such that row-sum normalization isrespected. The derivative r is mechanically zero if = = 0. So we set = 5, given that the parameter spaceis bounded and 2 [0 1). The other parameters that enter the derivative are estimated from a regression of on, with the full set of …xed e¤ects, and = 0.
38We also implemented an additional …ve particles. Particle 2: like Particle 1 but with size proportional to themagnitude of the derivatives conditional on ¡r being greater than 1; Particle 3: sets to non-zero all positiveelements of ¡r with equal weights; Particle 4: selects 5% highest values of ¡r , sets all others to zero, and non-zero gain equal weights; Particle 5: obtained from the Lasso regression of on the of others with penalization1; Particle 6: obtained from the Lasso regression of on the of others with penalization 1. In all cases,weights are rescaled by row-speci…c constants such that row-sum normalization is complied with. The remaining94 particles are uniformly randomly selected by the built-in MATLAB particle swarm algorithm.
39At each row, we pick the ¤ closest to the main diagonal of .40Note that the Elastic Net penalty 1
Pj j is invariant with respect to choices of if row-sum normalization
is imposed. Yet, the penalty a¤ects the initial selection of arguments in which is restricted to zero if thederivative of the objective function is smaller than 1 in absolute value.
46
where = [1 ], = , 0 = vec (¦00), and = [1 ]. If the number of individu-
als in the network is …xed and much smaller than data points available, 2 ¿ , equation
(18) can be estimated via ordinary least squares (OLS). Under suitable regularity conditions, the
OLS estimator = ( 0)¡1 0 is asymptotically distributed,
p ( ¡ 0)
¡! N
¡0 ¡1§¡1
¢
where ´ 1
0, ´ lim!1 , § ´ 1
00 and § ´ lim!1§ . The proof is
standard and omitted here. As noted above, in typical applications it is customary to row-sum
normalize matrix . If no individual is isolated, one obtains that, by equation (5),
¦0 = 0+ (00 + 0)1X
=1
¡10 0
=0 + 01¡ 0
(19)
where is the -length vector of ones. The last equality follows from the observation that, under
row-normalization of 0, = = , 0. Equation (19) implies that ¦0 has constant
row-sums, which implies that row-sum normalization is, in principle, testable. This suggests a
simple Wald statistic applied to the estimates of 0. Under the null hypothesis,
p
¡! N
¡0 ¡1§¡10
¢
where = [¡1 0 ;¡¡1 0 ]. The Wald statistic is = ()0 (¡1§¡1)¡1() »
2¡1 which is a convenient expression for testing row-sum normalization of 0. We also note that
the asymptotic distribution of can be immediately obtained by the Delta Method,
p ( ¡ 0)
¡! N
¡0r0¡1
¡1§¡1r
¢
where r is the gradient of with respect to . We note that the derivation of the Wald statistic
for testing the row-sum normalization and the asymptotic distribution of does not depend on
the OLS implementation, and can be easily adjusted for any estimator for which the asymptotic
distribution is known.
C Simulations
C.1 Set-Up
The simulations are based on two stylized random network structures, and two real world networks.
These networks vary in their size, complexity, and aggregate and node-level features. All four
47
networks are also sparse. The two stylized networks considered are:
(i) Erdos-Renyi network: we randomly pick exactly one element in each row of 0 and set that
element to 1. This is a random graph with in-degree equal to 1 for every individual (Erdos
and Renyi, 1960). Such a network could be observed in practice if connections are formed
independently of one another. With = 30, the resulting density of links is 345%.
(ii) Political party network: there are two parties, each with a party leader. The leader directly
a¤ects the behavior of half the party members. We assume that one party has twice the
number of members as the other. More speci…cally, we assume individuals = 1 3
are
a¢liated to Party A and are led by individual 1; individuals = 3+ 1 are a¢liated
to Party B and are led by individual 3+1. This di¤erence in party size allows us to evaluate
our ability to recover and identify central leaders, even in the smaller party. To test the
procedure further, we add one random link per row to represent ties that are not determined
by links to the Party leader. We simulate this network for various choices of . If is not
a multiple of three, we round 3
to the nearest integer. With = 30, this network has a
density of 517%.
(iii) Coleman’s (1964) high school friendship network survey: in 19578, students in a small high
school in Illinois were asked to name, “fellows that they go around with most often.” A link
was considered if the student nominated a peer in either survey wave. The full network has
= 73 nodes, of which 70 are non-isolated and so have at least one link to another student.
On average, students named just over …ve friendship peers. This network has density 758%.
Furthermore, the in-degree distribution shows that most individuals received a small number
of links, while a small number received many peer nominations.
(iv) Banerjee et al.’s (2013) village network survey: these authors conducted a census of house-
holds in 75 villages in rural Karnataka, India, and survey questions include several about
relationships with other households in the village. To begin with, we use social ties based
on family relations (later examining insurance networks). We focus on village 10 that is
comprised of = 77 households, and so similar in size to network (iii). In this village there
are 65 non-isolated households, with at least one family link to another household. This
network has density 507%.
For the stylized networks (i) and (ii), we …rst assess the performance of the estimator for a
…xed network size, = 30. We later show how performance varies with alternative network sizes.
We simulate the real-world networks (iii) and (iv) using non-isolated nodes in each (so = 70
and 65 respectively). As in Bramoullé et al., 2009, we exclude isolated nodes because they do not
conform with row-sum normalization.
48
Our result identi…es entries in 0 and so naturally recovers links of varying strength. Data
limitations often force researchers to postulate some ties to be weaker than others (say, based on
interaction frequency). This is in sharp contrast to our approach, that identi…es the continuous
strength of ties, 0, where 0 0 implies node in‡uences node .
To establish the performance of the estimator in capturing variation in link strength, we proceed
as follows for each network. First, for each node we randomly assign one of their links to have
value 0 = 7. As the underlying data generating process is assumed to allow for common time
e¤ects (), we then set the weight on all other links from the node to be equal and such that
row-sum normalization (A4’) is complied with.41 As we consider larger networks, we typically
expect them to have more non-zero entries in each row of 0, but row-sum normalization means
that each weaker link will be of lower value. This works against the detection of weaker links
using estimation methods involving penalization, because they impose small parameter estimates
shrink to zero.42 Finally, to aid exposition, we set a threshold value for link strength to distinguish
‘strong’ and ‘weak’ links. A strong (weak) link is de…ned as one for which 0 (·) 3.
Summary statistics for each network are presented in Panel A of Table A1. Following Jackson
et al. (2017), we consider the following network-wide statistics: number of edges, number of strong
and weak edges, number of reciprocated edges, clustering coe¢cient, number of components, and
the size of the maximal component. In addition, we report the standard deviation calculated
across elements of the diagonal of 20 . If this is zero, then the diagonal of 2
0 is either zero or
proportional to the vector of ones, and Assumption A5 would not be satis…ed. We can see that
for each case this statistic is well above zero.
Following Jackson et al. (2017), we consider the following node-level statistics: in- and out-
degree distribution (mean and standard deviation), and the three most central individuals. The
four networks di¤er in their size, complexity, and the relative importance of strong and weak ties.
For example, the Erdos-Renyi network only has strong ties, the political party network has twice
as many strong as weak ties. For the real world networks, the mean out-degree distributions are
higher so the majority of ties are weak, with the high school network having around 80% of its
edges being weak ties.
Panel data for each of the four simulations is generated as,
= ( ¡ 00)¡1( +0 + + ¤ + )
41For example, if in a given row of 0 there are two links, one will be randomly selected to be set to 7, and theother set to 3. If there are three links one is set to 7 and the other two set to each have weight 15 to maintainrow-sum normalization, and so on. For the Erdos-Renyi network, there are thus only strong ties as each node hasonly one link to another node.
42Caner and Zhang (2014) state that “local to zero coe¢cients should be larger than ¡12 to be di¤erentiated
from zero.”
49
where is a (scalar) time e¤ect and ¤ is a £ 1 vector of …xed e¤ects, drawn respectively from
(1 1) and ( £) distributions. We consider = f5 10 15 25 50 75 100 125 150g. The
true parameters are set to 0 = 3, 0 = 4 and 0 = 5 (thus satisfying Assumption A3). The
exogenous variable () and error term () are simulated as standard Gaussian, both generated
from (0 £) distributions. This is similar to variance terms set in other papers, e.g., Lee
(2004). We later conduct a series of robustness checks to evaluate the sensitivity of the simulations
to alternative parameters choices, and the presence of common- and individual-level shocks.
For each combination of parameters, we conduct 1 000 simulation runs. On the initial 50
runs, we choose penalization parameters that minimize the BIC criteria on a grid. This is
computationally intensive because it requires running the optimization procedure described in
Section 3.1 as many times as the number of points in the grid for .43 To reduce the computational
burden, we do so only in the initial 50 runs and consider these simulation runs as a calibration
of . For the remaining 950 iterations, the penalization parameter is set …xed at the median
computed over the calibration runs. This only worsens the performance of the estimator, since a
sub-optimal is chosen for the majority of the iterations.
C.2 Results
We evaluate the procedure over varying panel lengths (starting from short panels with = 5),
using the following metrics. Given our core contribution is to identify the social interactions matrix,
we …rst examine the proportion of true zero entries in 0 estimated as zeros, and the proportion
of true non-zero entries estimated as non-zeros. A global perspective of the proximity between
the true and estimated network can be inferred from their average absolute distance between
elements. This is the mean absolute deviation of and ¦ relative to their true values, de…ned
as ( ) = 1(¡1)
P6= j ¡ 0j and (¦) = 1
(¡1)
P6= j¦ ¡ ¦0j. The
closer these metrics are to zero, more of the elements in the true matrix are correctly estimated.
Finally, we evaluate the performance of the procedure using averaged estimates of the endogenous
and exogenous social e¤ect parameters, and . In keeping with the estimation strategy in our
empirical application, we report ‘post-Elastic Net’ estimates obtained after having estimated the
social interactions matrix by the Elastic Net GMM procedure. We use peers-of-peers’ covariates
from the estimated matrix as instrumental variables.
Each Panel in Figure A1 shows a di¤erent metric as we vary for each simulated network.
Panel A shows that for each network, the proportion of zero entries in 0 correctly estimated as
zeros is above 90% even when exploiting a small number of time periods ( = 5). The proportion
approaches 100% as grows. Panel B shows the proportion of non-zeros entries estimated as
non-zeros is also high for small . It is above 70% from = 5 for the Erdos-Renyi network, being
43In our simulations, we set the penalization grid to 1 = [0 025 05 10], ¤1 = [0 025 05 10] and 2 =
[0 025 05 10], resulting in 43 = 64 points per run.
50
at least 85% across networks for = 25, and increasing as grows.
Panels C and D show that for each simulated network, the mean absolute deviation between
estimated and true networks for and ¦ falls quickly with and is close zero for large sample
sizes. Finally, Panels E and F show that biases in the endogenous and exogenous social e¤ects
parameters, and , also fall quickly in (we do not report the bias in since it is close to
zero for all ). The fact that biases are not zero is as expected for small , being analogous to
well-known results for autoregressive time series models.
Figure A2 provides a visual representation of the simulated and actual networks under = 100
time periods. The network size is set to = 30 in the two stylized networks, = 70 for the high
school network, and = 65 for the village household network. In comparing the simulated and
true network, Figure A2 distinguishes between three types of edges: kept edges, added edges and
removed edges. Kept edges are depicted in blue: these links are estimated as non-zero in at least
5% of the iterations and are also non-zero in the true network. Added edges are depicted in green:
these links are estimated as non-zero in at least 5% of the iterations but the edge is zero in the
true network. Removed edges are depicted in red: these links are estimated as zero in at least 5%
of the iterations but are non-zero in the true network. Figure A2 further distinguishes between
strong and weak links: strong links are shown in solid edges (0 3), and weak links are shown
as dashed edges.
Consider …rst Panel A of Figure A2, comparing the simulated and true Erdos-Renyi network.
All zero and all non-zero links are correctly estimated. All links are thus recovered and no edges are
added to the true network (all edges are in blue). For the political party network, Panel B shows
that all strong edges are correctly estimated (it also highlights the party leader nodes). However,
around half the weak edges are recovered (blue dashed edges) with the others being missed (red
dashed edges). As discussed above, this is not surprising given that shrinkage estimators force
small non-zero parameters to zero. Hence, larger is needed to achieve similar performance as
in the other simulated networks in terms of detecting weak links. Again, we never estimate any
added edges (no edges are green).
For the more complex and larger real-world networks, Panel C shows that in the high school
network, strong edges are all recovered. However, around half the weak edges are missing (red
dashed edges) and there are a relatively small number of added edges (green edges): these amount
to 87 edges, or approximately 19% of the 4 534 zero entries in the true high-school network. A
similar pattern of results is seen in the village network in Panel D: strong edges are all recovered,
and here the majority of weak edges are also recovered. A relatively small share of overall edges
are added or missed.
Panel B of Table A1 compares the network- and node-level statistics calculated from the re-
covered social interactions matrix to those in Panel A from the true interactions matrix 0.
As Figure A2 showed, the random Erdos-Renyi network is perfectly recovered. For the political
party network, the number of recovered edges is slightly lower than the true network (38 vs. 45).
51
This is driven by weak edges: while all the strong edges are recovered (30 out of 30), not all the
weak ones are (8 vs. 15). On node-level statistics, the mean of the in- and out-degree distributions
are slightly lower in the recovered network, the clustering coe¢cient is exactly recovered, and all
three nodes with the highest out-degree are correctly captured (nodes 1, 11 and 28), that includes
both party leaders (individuals 1 and 11).
Performance in the two real world networks is also encouraging. In the high school network,
all strong edges are correctly recovered, as are the majority of weak edges. However, as already
noted in Figure A2, because weak edges are not well estimated in the high school network, the
average in- and out- degrees are smaller in the recovered network relative to the true network. We
recover two out of the three individuals with the highest out-degree (nodes 21 and 69). Finally,
in the village network, all strong edges are recovered, the majority of weak edges are recovered,
the clustering coe¢cients are similar across recovered and true networks (134 vs. 141) and we
recover two out of the three households with the highest out-degree (nodes 16, 35, and 57).
C.3 Robustness
Table A2 presents results for the recovered stylized networks under varying network sizes, =
f15 30 50g. Di¤erences between the true and estimated networks are fairly constant as in-
creases: even for small = 15 a large proportion of zeros and non-zeros are correctly estimated.
In all cases, biases in and decrease with larger . We also conduct a counterpart robustness
check for one of the real work networks. More precisely, we use the savings and insurance networks
between households in villages identi…ed in Banerjee et al. (2013), that are generally larger than
family networks focused on so far. Table A3 shows descriptive statistics on this true village net-
work (Panel A) and the recovered network (Panel B). Relative to the family network, the savings
and insurance network has many more edges, a greater proportion of weak edges, is less clustered,
with nodes having a higher degree distribution. Despite these di¤erences, the recovered network
retains good accuracy on many dimensions: 78% of all edges are recovered, the recovered clustering
coe¢cient is 058 (relative to an actual coe¢cient of 073) and the three nodes with the highest
out-degree are all still identi…ed.
Table A4 conducts robustness checks on the sensitivity of the estimates to parameters choices.
We consider true parameters 0 = f1 3 7 9g, 0 = f3 7g, 0 = f0 8g. We also introduce a
common shock in the disturbance variance-covariance matrix by varying in,
»
0
BBBB@0
2
66664
1 ¢ ¢ ¢
1 ¢ ¢ ¢ ...
.... . .
...
¢ ¢ ¢ 1
3
77775
1
CCCCA
where we consider = f3 5 8 1g. We …nd the procedure to be robust to the true values of
52
0, 0, 0, and . For 0 = 0, performance is slightly worse. This is expected as the exogenous
variation from no longer a¤ects directly.
We next probe the procedure by richening up the structure of shocks across nodes. First, we
introduce a common shock correlated with covariates . To do so, we take from a Gaussian
distribution with mean 05 and, as before, variance 1. Second, we implement a version where
the shock is constant over time but varies at the individual level. In this case, the mean of is
given by 05¤. Third, we implement a version mixing the two types of shocks, with the mean of
given by 05¤+05. In each case, we simulate based on the Erdos-Renyi network as the true
0. The results are shown in Figure A3: this shows that for each of the six performance metrics
considered, the procedure is highly robust to these richer structures of shocks across individuals
and time periods.
The …nal robustness check demonstrates the gains from using the Adaptive Elastic Net GMM
estimator over alternative estimators. Table A5 shows simulation results using Adaptive Lasso
estimates of the interaction matrix ¦0, so estimating and penalizing the reduced-form. The Ad-
aptive Lasso estimator performs relatively worse: the mean absolute deviation between and 0
is often two to three times larger than the corresponding Adaptive Elastic Net estimates. Appendix
Table A6 then shows the performance of the procedure based on OLS estimates of ¦0. Given OLS
requires ¿ , we use a time dimension ten times larger, = f500 1000 1500g, and still …nd
a deterioration in performance compared to the Adaptive Elastic Net GMM estimator.
Taken together, these robustness checks suggest the Adaptive Elastic Net GMM estimator is
preferred over Adaptive Lasso and OLS estimators. This procedure does well in recovering true
network structures, and a fortiori, network- and node-level statistics. It does so in networks that
vary in size and complexity, and as the underlying social interactions model varies in the strength
of endogenous and exogenous social e¤ects, and the structure of shocks.
D Application: Counterfactuals
We consider a scenario in which California exogenously increases the change in its taxes per capita,
so ¢ corresponds to an increase of 10%. We measure the di¤erential change in equilibrium state
taxes in state under the two network structures using the following statistic:
¨ = log(¢ j)¡ log(¢ j) (20)
so that positive (negative) values imply taxes being higher (or lower) under than .44
Figure A4 graphs ¨ for each mainland US state (including for California itself, the origin of
the shock). A wide discrepancy between the equilibrium state tax rates predicted under
44We calculate the counterfactual at 2 = 608, the endogenous e¤ect parameter estimated in our preferredspeci…cation, Column 7 of Table 2.
53
relative to : across states ¨ varies from ¡303 to 961. Only in one state is ¨ close to
zero. Table A9 summarizes the general equilibrium e¤ects under both structures. We see that
average tax rate increases are 74% higher under . The dispersion of tax rates across states
also increases dramatically under . Finally, assuming interactions across states are based
solely on geographic neighbors, we miss the fact that many states will have relatively small tax
increases.
54
Table 1: Geographic Neighbors
Dependent variable: Change in per capita income and corporate taxes
Coefficient estimates, standard errors in parentheses
(1) OLS (2) 2SLS (3) OLS (4) 2SLS
Geographic Neighbors' Tax Change (t - [t-2]) .375*** .868*** .271*** .642***
(.120) (.273) (.075) (.152)
Period 1962-1988 1962-1988 1962-2015 1962-2015
First Stage (F-stat) 6.267 27.320
Controls Yes Yes Yes Yes
State and Year Fixed Effects Yes Yes Yes Yes
Observations 1,296 1,248 2,592 2,544
Extended SampleBesley and Case (1995) Sample
Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. In all specifications, a pair of states are considered neighbors if they share a geographic border. The
sample covers 48 mainland US states. In Columns 1 and 2 the sample runs from 1962 to 1988 (as in Besley and Case (1995)). In Columns 3 and 4 the sample isextended to run from 1962 to 2015. The dependent variable is the change in state i's total taxes per capita in year t. OLS regressions estimates are shown in Columns 1and 3. Columns 2 and 4 show 2SLS regressions where each geographic neighbors' tax change is instrumented by lagged neighbor's state income per capita andunemployment rate. At the foot of Columns 2 and 4 we report the p-value on the F-statistic from the first stage of the null hypothesis that instruments are jointly equal tozero. All regressions control for state i’s income per capita in 1982 US dollars, state i’s unemployment rate, the proportion of young (aged 5-17) and elderly (aged 65+) instate i’s population, and the state governor's age. All specifications include state and time fixed effects. With the exception of governor's age, all variables are differencedbetween period t and period t-2. Robust standard errors are reported in parentheses.
Table 2: Economic Neighbors
Dependent variable: Change in per capita income and corporate taxes
Coefficient estimates, standard errors in parentheses
(1) Initial (2) OLS
(3) 2SLS: IVs are
Characteristics of
Neighbors
(4) Initial (5) OLS
(6) 2SLS: IVs are
Characteristics of
Neighbors
(7) 2SLS: IVs are
Characteristics of
Neighbors-of
Neighbors
Economic Neighbors' Tax Change (t - [t-2]) .886 .378*** .641*** .645 .145** .332* .608***
(.061) (.060) (.072) (.199) (.220)
Period
First Stage (F-stat) 19.353 9.571 10.480
Controls Yes Yes Yes Yes Yes Yes Yes
State and Year Fixed Effects Yes Yes Yes Yes Yes Yes Yes
Observations 2,952 2,952 2,544 2,952 2,952 2,544 2,592
Exogenous Social Effects
1962-20151962-2015
No Exogenous Social Effects
Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The sample covers 48 mainland US states running from 1962 to 2015. The dependent variable is the change in state i's total
taxes per capita in year t. We allow for exogenous social effects in Columns 4 to 7. In subsequent OLS and IV regressions, the economic neighbors' effect is calculated as the weightedaverage of economic neighbors' variables. OLS regressions estimates are shown in Column 2, 3 and 5. Column 3 and 6 show the 2SLS regression where each geographic neighbors' tax
change is instrumented by lagged neighbor's state income per capita and unemployment rate. Column 7 shows a 2SLS regression where each geographic neighbors' tax change is
instrumented by lagged neighbor-of-neighbor's state income per capita and unemployment rate. At the foot of Columns 3, 6 and 7 we report the p-value on the F-statistic from the firststage of the null hypothesis that instruments are jointly equal to zero. All regressions control for state i’s income per capita in 1982 US dollars, state i’s unemployment rate, the proportion ofyoung (aged 5-17) and elderly (aged 65+) in state i’s population, and the state governor's age. All specifications include state and time fixed effects. With the exception of governor's age,
all variables are differenced between period t and period t-2. Robust standard errors are reported in parentheses.
Table 3: Geographic Versus Economic Neighbor Networks
Geographic NeighborNetwork
Economic NeighborNetwork
Number of Edges 214 144
Edges in Both Networks 79 79
Edges in W-geo only 135
Edges in W-econ only 65
Clustering .1936 .0259
Reciprocated Edges 100% 29.17%
out-degree 3.000 (1.185)
in-degree 3.000 (2.073)4.458 (1.597)
Degree Distribution Across Nodes (states)
Notes: This compares statistics derived from the geographic network of US states to those from the estimated economic
network among US states. The number of edges, edges in both networks, edges in W-geo only, edges in W-econ onlycounts the number of edges in those categories. Reciprocated edges is the frequency of in-edges that are reciprocated byout-edges (by construction, this is 100% for geographic networks). The clustering coefficient is the frequency of thenumber of fully connected triplets over the total number of triplets. The degree distribution across nodes counts theaverage number of connections (standard deviation in parentheses): we show this separately for in-degree and out-degree(by construction, these are identical for geographic networks).
Table 4: Predicting Links to Economic Neighbors
Columns 1-7: Linear Probability Model; Column 8: Tobit
Dependent variable (Cols 1-7): =1 if Economic Link Between States Identified
Dependent variable (Col 8): =Weighted Link Between States
Coefficient estimates, standard errors in parentheses
Economic and
Demographic
Homophyly
Labor
Mobility
Political
Homophyly
Tax
Havens
Tobit, Partial
Avg Effects
(1) (2) (3) (4) (5) (6) (7) (8)
Geographic Neighbor .699*** .701*** .701*** .698*** .698*** .697*** .068***
(.030) (.032) (.030) (.031) (.031) (.031) (.006)
Distance -.453*** -.008
(.033) (.024)
Distance sq. .0949*** .003
(.007) (.006)
GDP Homophyly 2.409** 2.369* 2.296* 1.046 .322
(1.183) (1.186) (1.193) (1.150) (.302)
Demographic Homophyly .222 .235 .241 .256 .077
(.226) (.226) (.228) (.225) (.067)
Net Migration .044* .044* -0.032 0.001
(.025) (.025) (.025) (.002)
Political Homophyly -.057 -.083** -.025*
(.042) (.042) (.014)
Tax Haven Sender .107*** .021***
(.024) (.005)
Adjusted R-squared 0.427 0.152 0.427 0.428 0.429 0.429 0.440 -
Observations 2,256 2,256 2,256 2,256 2,256 2,256 2,256 2,256
Geography
Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The specifications in Columns 1-7 are cross-sectional linear probabilities models
where the dependent variable is equal to 1 if two states are linked, and zero otherwise. In Column 8 the dependent variable is the weighted linkbetween states. Column 8 reports the partial average effects from a Tobit model. A pair of states is considered a first-degree geographic neighbor ifthey share a border. Distance and distance squared are calculated from the centroids of states' capital cities. GDP homophyly is the absolutedifference of states' GDP per capita. Demographic homophyly is the absolute difference of share of young (aged 5-17) plus the absolute differenceof the share of elderly in states' population (aged 65+). Net migration based on individuals tax returns (Source: Internal Revenue Service,https://www.irs.gov/statistics/soi-tax-stats-migration-data). Political homophyly is equal to one if a pair of states have governors of same party atgiven year. Nevada, Delaware, Montana, South Dakota, Wyoming and New York are considered tax haven states. Time averages are taken for allexplanatory variables. Robust standard errors in parentheses.
Table 5: Gubernatorial Term Limits
Dependent variable: Change in per capital income and corporate taxes
Coefficient estimates, standard errors in parentheses
IVs: Characteristics of Neighbors-of-Neighbors
(1) OLS (2) 2SLS (3) OLS (4) 2SLS (5) OLS (6) 2SLS
Economic Neighbors' tax change (t - [t-2]) .145** .608*** .016 .937* .182** .543**
(.072) (.220) (.105) (.534) (.084) (.237)
Period
First Stage (F-stat) 10.480 2.835 10.120
Controls Yes Yes Yes Yes Yes Yes
State and Year Fixed Effects Yes Yes Yes Yes Yes Yes
Observations 2,592 2,592 640 640 1,917 1,917
Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The sample in Columns 1 and 2 covers 48 mainland US states running from 1962 to 2015. In Columns 3 and 4 we use
the subsample of state-years in which the governor that faced term limits in the subsequent gubernatorial election. In Columns 5 and 6 we use the subsample of state-years in which
the governor did not face term limits in the subsequent gubernatorial election, and so could run for reelection. The dependent variable is the change in state i's total taxes per capita
in year t. We first estimate our procedure which outputs parameters and the network of economic neighbors. We penalize geographic neighbors throughout and also allow for
exogenous social effects. OLS regressions estimates are shown in Columns 1, 3 and 5. Columns 2, 4 and 6 show a 2SLS regression where each geographic neighbors' tax change
is instrumented by lagged neighbor-of-neighbor's state income per capita and unemployment rate. At the foot of Columns 2, 4 and 6 we report the p-value on the F-statistic from the
first stage of the null hypothesis that instruments are jointly equal to zero. All regressions control for state i’s income per capita in 1982 US dollars, state i’s unemployment rate, the
proportion of young (aged 5-17) and elderly (aged 65+) in state i’s population, and the state governor's age. All specifications include state and time fixed effects. With the exception
of governor's age, all variables are differenced between period t and period t-2. Robust standard errors are reported in parentheses.
1962-2015 1962-2015 1962-2015
All GovernorsGovernor Cannot Run for Re-
election
Governor Can Run for Re-
election
Geographic network edges
Removed (geographic) edges in economic network
New edges added in economic networks
Figure 1B: Network Graph of US States, Identified Economic Neighbors
Notes: Figure 3B represents the continental United States (N=48). The economic network is derived from our preferred specification, where we penalize geographic neighbors to states, and
allow for exogenous social effects. A blue edge is drawn between a pair of states if they are geographic neighbors and were estimated as connected. A red edge is drawn between a pair of states
if they are geographic neighbors but were not estimated as connected. A green edge is drawn between a pair of states if they are not geographic neighbors and were estimated connected. The
left hand side graph just shows read and blue edges. The right hand side shows all three types of edges. State abbreviations are as used by US Post Office (http://about.usps.com/who-we-
are/postal-history/state-abbreviations.pdf).
Notes: Figure 3A represents the continental US states (N=48). An edge is drawn between a pair of states if they share a geographic border. State abbreviations are as used by US Post Office
(http://about.usps.com/who-we-are/postal-history/state-abbreviations.pdf).
Figure 1A: Network Graph of US States, Geographic Neighbors
Geographic network edgesRemoved (geographic) edges in economic networkNew edges added in economic networks
Panel B: In-network for California
Figure 2: Strength of Ties and Reciprocity
Panel C: Out-network for California
Panel A: Histogram of Strength of Ties, Conditional on W0,ij>0
Notes: Panel A is the histogram of ties in the economic network, conditional on non-zero ties. Panels B and C show the in-network and out-network of California as derived
from our preferred specification, where we penalize geographic neighbors to states, and allow for exogenous social effects. The in-network are the states that determine tax
setting in California. The out-network is the states in which taxes are set in direct response to those in California. A blue edge is drawn between a pair of states if they are
geographic neighbors and were estimated as connected. A red edge is drawn between a pair of states if they are geographic neighbors but were not estimated as connected.
A green edge is drawn between a pair of states if they are not geographic neighbors and were estimated connected. State abbreviations are as used by US Post Office
(http://about.usps.com/who-we-are/postal-history/state-abbreviations.pdf).
Figure 3: In- and Out-degree Distribution
Panel A: In-degree distribution Panel B: Out-degree distribution
Notes: In-degree distribution (Panel A) and out-degree distribution (Panel B). Distribution calculated from geographic neighbors' network (W-geo) in blue. Distribution calculated from economic neighbor's
network in (W-econ) in red. State abbreviations are as used by US Post Office (http://about.usps.com/who-we-are/postal-history/state-abbreviations.pdf).
A. % of zeros B. % of non-zeros
C. Mean Absolute Deviation of D. Mean Absolute Deviation of
E. Endogenous Social Effect, F. Exogenous Social Effect
Notes: These simulation results are based on the Banerjee et al. (2013) village network, using the Adaptive Elastic Net GMM
algorithm, with penalization parameters chosen by BIC, under various assumptions about knowledge of the true network and
time periods T=25, 50, 100, 125 and 150. The “Village" case refers to the simulation implemented without knowledge about the
true network. "Village (top 3)" refers to the case where all connections of the three households with highest out-degrees are
assumed to be known. "Village (top 5)" and "Village (top 103)" are analogously defined. In all cases, 1000 Monte Carlo
iterations were performed. The true parameters are rho-0=.3, beta-0=.4 and gamma-0=.5. In Panel A, the % of zeroes refers to
the proportion of true zero elements in the social interaction matrix that are estimated as smaller than .05. In Panel B, the % of
non-zeros refers to the proportion of true elements greater than .3 in the social interaction matrix that are estimated as non-
zeros. In Panels C and D, the Mean Absolute Deviations are the mean absolute error of the estimated network compared to the
true network for the social interaction matrix W and the reduced form matrix respectively. In Panels E and F, the true parameter
values are marked in the horizontal red lines. The recovered parameter are the estimated parameters averaged across
iterations. All specifications include time and node fixed effects.
Figure 4: Simulation Results, Adaptive Elastic Net GMMPartial Knowledge of W0
Π
ොߩ ොߛ
T=25 50 75 100 125 1500.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
T=25 50 75 100 125 1500
0.02
0.04
0.06
0.08
0.1
0.12
T=25 50 75 100 125 1500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
T=25 50 75 100 125 1500.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
T=25 50 75 100 125 1500
0.02
0.04
0.06
0.08
0.1
0.12
T=25 50 75 100 125 1500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Panel B: Econ Network, 1989-2015 subsample
Panel C: Geographic Versus Economic Neighbor NetworksGeographic
Neighbor
Network
Economic Neighbor
Network (Full Sample)
Economic Neighbor
Network (1962-1988)
Economic Neighbor
Network (1989-2015)
Number of Edges 214 144 197 185
Edges in both W-geo and W-econ 79 108 86
Edges in W-econ only 65 89 99
Edges in W-geo only 135 106 128
Clustering .1936 .0259 .0389 .0394
Reciprocated Edges 100% 29.17% 35.53% 23.78%
out-degree 3.000 (1.185) 4.104 (1.704) 3.854 (1.473)
in-degree 3.000 (2.073) 4.104 (2.707) 3.854 (2.518)
Notes: This compares statistics derived from the geographic network of US states to those from the estimated economic network among US states, for the three samples ("Full Sample", 1962-
2015; 1962-1988; 1989-2015). The number of edges, edges in both networks, edges in W-geo only, edges in W-econ only counts the number of edges in those categories. Numbers are relative
to the W-geo network in the first column. Reciprocated edges is the frequency of in-edges that are reciprocated by out-edges (by construction, this is 100% for geographic networks). The
clustering coefficient is the frequency of the number of fully connected triplets over the total number of triplets. The degree distribution across nodes counts the average number of connections
(standard deviation in parentheses): we show this separately for in-degree and out-degree (by construction, these are identical for geographic networks).
Panel A: Econ Network, 1962-1988 subsample
Figure 5: Network Graph of US States, Identified Economic Neighbors by Subsamples
Notes: Panels A and B represents the continental United States (N=48). Panel A is estimated withe 1962-1988 subsample. Panel B is estimated with the 1989-2015 subsample. The economic
network is derived from our preferred specification, where we penalize geographic neighbors to states, and allow for exogenous social effects. A blue edge is drawn between a pair of states if they
are geographic neighbors and were estimated as connected. A red edge is drawn between a pair of states if they are geographic neighbors but were not estimated as connected. A green edge is
drawn between a pair of states if they are not geographic neighbors and were estimated connected. The left hand side graph just shows read and blue edges. The right hand side shows all three
types of edges. State abbreviations are as used by US Post Office (http://about.usps.com/who-we-are/postal-history/state-abbreviations.pdf).
Degree Distribution Across Nodes (states)
4.458 (1.597)
Erdos-Renyi Political Party High school VillageColeman (1964) Banerjee et al. (2013)
Number of nodes 30 30 70 65
(a) Network-wide statistics
Number of edges 30 45 366 240
Number of strong edges 30 30 70 65
Number of weak edges 0 15 296 175
Number of reciprocated edges 2 2 184 240
Clustering coefficient - .000 .120 .141
Number of components 12 11 3 3
Size of maximal component 10 16 68 51
Standard deviation of the
diagonal of squared W
(b) Node-level statistics
In-degree distribution 1.00 (0.00) 1.50 (.509) 5.23 (2.04) 3.69 (2.35)
Out-degree distribution 1.00 (1.05) 1.50 (2.49) 5.23 (3.64) 3.69 (2.35)
Nodes with highest out-degree { 7, 11, 26 } { 1, 11 , 28 } { 21, 22, 69 } { 16, 35, 57 }
(a) Network-wide statistics
Number of edges 30 38 210 194
Number of strong edges 30 30 70 68
Number of weak edges 0 8 140 126
Number of reciprocated edges 2 2 184 170
Clustering coefficient - .000 .162 .134
Number of components 12 11 1 4
Size of maximal component 10 14 70 48
(b) Node-level statistics
In-degree distribution 1.00 (0.00) 1.27 (.450) 3.00 (1.18) 2.99 (1.29)
Out-degree distribution 1.00 (1.05) 1.27 (1.76) 3.00 (1.02) 2.99 (1.15)
Nodes with highest out-degree { 7, 11, 26 } { 1, 11, 28 } { 21, 48, 69 } { 16, 35, 57 }
A. True Networks
Table A1: True and Recovered Networks
Notes: Panel A refers to the true networks. Panel B refers to the recovered networks. In each Panel, the summary statistics are divided into
network-wide and node-level statistics. Strong edges are defined as those with strength greater than or equal to .3. For the in-degree and out-
degree distribution, the mean is shown and the standard deviation is in parentheses. The nodes with the highest out-degree are those with the
greatest influence on others, and are calculated as the column-sum of the social interaction matrix. The recovered networks statistic are calculated
over the average network across simulations with T=100.
.254 .254 .167 .239
B. Recovered Networks
Table A2: Simulation Results, Adaptive Elastic Net GMM, Alternative Network Sizes
T=50 100 150 T=50 100 150 T=50 100 150 T=50 100 150 T=50 100 150 T=50 100 150
% True Zeroes .954 .976 .973 .957 .992 .997 .947 .965 .979 .950 .977 .979 .943 .978 .989 .943 .961 .976
(.033) (.031) (.031) (.009) (.004) (.002) (.007) (.006) (.005) (.031) (.025) (.024) (.009) (.006) (.005) (.007) (.006) (.006)
% True Non-Zeroes .899 .919 .924 .958 .960 .962 .977 .978 .977 .914 .925 .932 .930 .954 .960 .970 .977 .977
(.051) (.032) (.026) (.017) (.016) (.011) (.007) (.007) (.008) (.038) (.022) (.009) (.037) (.023) (.016) (.015) (.009) (.007)
.021 .007 .004 .017 .003 .001 .020 .011 .006 .032 .019 .017 .028 .014 .010 .023 .014 .009
(.014) (.011) (.008) (.004) (.001) (.001) (.002) (.001) (.001) (.011) (.007) (.004) (.004) (.002) (.002) (.002) (.001) (.001)
.025 .012 .009 .020 .004 .003 .054 .016 .007 .034 .021 .019 .028 .014 .012 .065 .019 .009
(.012) (.008) (.006) (.004) (.002) (.001) (.018) (.004) (.002) (.009) (.006) (.003) (.004) (.002) (.001) (.028) (.004) (.001)
.262 .270 .276 .286 .286 .283 .667 .398 .270 .235 .245 .241 .209 .228 .223 .700 .383 .242
(.069) (.044) (.038) (.079) (.026) (.022) (.079) (.050) (.050) (.078) (.049) (.038) (.084) (.035) (.029) (.100) (.069) (.051)
.403 .399 .400 .405 .400 .400 .380 .399 .401 .405 .400 .397 .404 .399 .398 .380 .400 .401
(.039) (.028) (.022) (.028) (.018) (.015) (.025) (.015) (.012) (.040) (.028) (.022) (.029) (.019) (.015) (.024) (.015) (.012)
.577 .521 .507 .670 .516 .498 .748 .686 .595 .550 .481 .459 .669 .501 .463 .686 .666 .566
(.094) (.059) (.046) (.054) (.022) (.018) (.118) (.056) (.028) (.093) (.060) (.050) (.060) (.027) (.021) (.128) (.063) (.032)
Notes: These simulation results are based on the Adaptive Elastic Net GMM algorithm, with penalization parameters chosen by BIC, under various true networks, network sizes and
time periods T=50, 100 and 150. In all cases, 1000 Monte Carlo iterations were performed. The true parameters are rho-0=.3, beta-0=.4 and gamma-0=.5. The % of true zeroes refers
to the proportion of true zero elements in the social interaction matrix that are estimated as smaller than .05. The % of true non-zeroes refers to the proportion of true elements greater
than .3 in the social interaction matrix that are estimated as non-zeros. The Mean Absolute Deviations are the mean absolute error of the estimated network compared to the true
network for the social interaction matrix W and the reduced form matrix respectively. The recovered parameter are the estimated parameters averaged across iterations. All
specifications include time and node fixed effects. Standard errors across iterations are in parentheses.
N = 15 N = 30 N = 50
A. Erdos-Renyi B. Political party
N = 15 N = 30 N = 50
ࢰ
ෝ
ࢼ
ෝࢽ
Table A3: True and Recovered Village Networks
Village VillageFamily Savings and Insurance
Banerjee et al. (2013) Banerjee et al. (2013)
Number of nodes 65 65
(a) Network-wide statistics
Number of edges 240 343
Number of strong edges 65 47
Number of weak edges 175 296
Number of reciprocated edges 240 340
Clustering coefficient .141 .073
Number of components 3 6
Size of maximal component 51 62
Standard deviation of the
diagonal of squared W
(b) Node-level statistics
In-degree distribution 3.69 (2.35) 4.90 (3.42)
Out-degree distribution 3.69 (2.35) 4.90 (3.43)
Nodes with highest out-degree { 16, 35, 57 } { 16, 35, 55 }
(a) Network-wide statistics
Number of edges 194 269
Number of strong edges 68 65
Number of weak edges 126 204
Number of reciprocated edges 170 250
Clustering coefficient .134 .058
Number of components 4 4
Size of maximal component 48 62
(b) Node-level statistics
In-degree distribution 2.99 (1.29) 3.84 (1.90)
Out-degree distribution 2.99 (1.15) 3.84 (1.98)
Nodes with highest out-degree { 16, 35, 57 } { 16, 35, 55 }
.159
Notes: Panel A refers to the true networks. Panel B refers to the recovered networks. In each Panel, thesummary statistics are divided into network-wide and node-level statistics. Strong edges are defined as thosewith strength greater than or equal to .3. For the in-degree and out-degree distribution, the mean is shown andthe standard deviation is in parentheses. The nodes with the highest out-degree are those with the greatestinfluence on others, and are calculated as the column-sum of the social interaction matrix. The recoverednetworks statistic are calculated over the average network across simulations with T=100.
.239
A. True Networks
B. Recovered Networks
Table A4: Simulation Results, Adaptive Elastic Net GMM, Alternative Parameters
.1 .5 .7 .9 .0 .8 .3 .7 .3 .5 .8 1.0 .1 .5 .7 .9 .0 .8 .3 .7 .3 .5 .8 1.0
% True Zeroes .986 .994 .991 .979 .986 .987 .974 .997 .996 .997 .997 .997 .971 .983 .982 .966 .978 .977 .959 .986 .985 .992 .996 .997
(.005) (.004) (.004) (.005) (.005) (.007) (.006) (.002) (.002) (.003) (.002) (.002) (.008) (.005) (.006) (.007) (.008) (.007) (.007) (.005) (.004) (.004) (.004) (.003)
% True Non-Zeroes .951 .963 .963 .956 .806 .967 .961 .961 .963 .962 .961 .959 .772 .813 .836 .864 .469 .856 .741 .834 .803 .808 .816 .832
(.028) (.011) (.011) (.017) (.096) .000 (.015) (.016) (.012) (.013) (.013) (.018) (.045) (.036) (.034) (.031) (.147) (.035) (.044) (.036) (.037) (.035) (.034) (.027)
.005 .002 .003 .007 .014 .004 .011 .001 .001 .001 .000 .000 .017 .012 .012 .018 .029 .012 .023 .010 .011 .009 .008 .007
(.002) (.001) (.001) (.002) (.005) (.002) (.003) (.001) (.001) (.001) (.001) (.001) (.003) (.002) (.002) (.003) (.007) (.002) (.003) (.002) (.002) (.001) (.001) (.001)
.005 .006 .013 .077 .013 .007 .009 .004 .003 .003 .002 .002 .011 .021 .043 .208 .021 .017 .016 .016 .013 .011 .010 .009
(.002) (.002) (.005) (.051) (.004) (.003) (.002) (.002) (.001) (.001) (.001) (.001) (.002) (.002) (.004) (.026) (.004) (.003) (.002) (.002) (.002) (.001) (.001) (.001)
.081 .487 .709 .917 .318 .244 .287 .286 .285 .285 .288 .292 .046 .403 .598 .847 .298 .171 .221 .225 .224 .222 .218 .218
(.034) (.027) (.025) (.012) (.047) (.076) (.036) (.023) (.023) (.020) (.017) (.009) (.038) (.031) (.040) (.040) (.071) (.067) (.048) (.030) (.030) (.028) (.025) (.018)
.401 .399 .397 .364 (.008) .801 .402 .399 .401 .400 .400 .400 .400 .399 .398 .380 -.015 .801 .402 .400 .399 .398 .398 .398
(.018) (.019) (.019) (.023) (.025) (.018) (.019) (.019) (.016) (.012) (.009) (.001) (.019) (.020) (.020) (.026) (.029) (.020) (.020) (.019) (.016) (.014) (.009) (.003)
.532 .512 .517 .509 .432 .552 .376 .698 .503 .495 .493 .492 .519 .489 .488 .469 .305 .548 .391 .638 .474 .454 .440 .431
(.023) (.022) (.024) (.028) (.070) (.033) (.027) (.021) (.018) (.015) (.012) (.007) (.030) (.027) (.031) (.050) (.117) (.038) (.036) (.025) (.023) (.019) (.015) (.010)
B. Political party
Notes: These simulation results are based on the Adaptive Elastic Net GMM algorithm, with penalization parameters chosen by BIC, under various true networks, network sizes, time periods T=100 and parameter values. In all cases,
1000 Monte Carlo iterations were performed. The % of true zeroes refers to the proportion of true zero elements in the social interaction matrix that are estimated as smaller than .05. The % of true non-zeroes refers to the proportion of
true elements greater than .3 in the social interaction matrix that are estimated as non-zeros. The Mean Absolute Deviations are the mean absolute error of the estimated network compared to the true network for the social interaction
matrix W and the reduced form matrix respectively. The recovered parameter are the estimated parameters averaged across iterations. All specifications include time and node fixed effects. Standard errors across iterations are in
parentheses.
A. Erdos-Renyi
ࢼ ࢽ
ࢰ
ෝ
ࢼ
ෝࢽ
ࢼ ࢽ
T=50 100 150 T=50 100 150 T=50 100 150 T=50 100 150 T=50 100 150 T=50 100 150
% True Zeroes .728 .798 .842 .737 .848 .878 .845 .846 .913 .730 .799 .840 .737 .849 .880 .846 .846 .921
(.023) (.024) (.022) (.012) (.008) (.008) (.005) (.006) (.027) (.024) (.025) (.023) (.013) (.009) (.009) (.006) (.006) (.008)
% True Non-Zeroes .996 1.000 1.000 .497 1.000 1.000 .494 .496 .930 .992 1.000 1.000 .507 1.000 1.000 .498 .501 .996
(.039) (.000) (.000) (.096) (.001) (.000) (.071) (.074) (.177) (.023) (.002) (.000) (.105) (.004) (.001) (.074) (.076) (.042)
.071 .050 .039 .064 .039 .033 .039 .039 .027 .070 .050 .039 .063 .038 .032 .039 .038 .025
(.005) (.005) (.004) (.002) (.002) (.002) (.000) (.002) (.005) (.005) (.004) (.004) (.002) (.002) (.002) (.000) (.002) (.001)
.092 .056 .042 .084 .049 .036 .071 .045 .032 .093 .057 .044 .084 .050 .036 .071 .045 .033
(.008) (.005) (.004) (.005) (.003) (.002) (.004) (.003) (.002) (.008) (.005) (.003) (.005) (.003) (.002) (.004) (.002) (.002)
.962 .815 .535 .998 .988 .965 1.000 .993 .995 .970 .783 .512 .998 .993 .979 1.000 .994 .995
(.112) (.186) (.231) (.040) (.066) (.095) (.000) (.075) (.054) (.106) (.199) (.236) (.041) (.050) (.077) (.000) (.069) (.054)
.131 .285 .330 .000 .158 .254 .000 .000 .121 .144 .292 .336 .000 .177 .259 .000 .000 .167
(.079) (.049) (.038) (.000) (.053) (.030) (.000) (.000) (.066) (.081) (.049) (.039) (.000) (.047) (.030) (.000) (.000) (.044)
.996 .998 .968 .000 1.000 .999 .000 .000 .863 1.000 .995 .942 .000 1.000 .999 .000 .000 .992
(.063) (.015) (.051) (.000) (.000) (.007) (.000) (.000) (.344) (.000) (.023) (.066) (.000) (.000) (.014) (.000) (.000) (.089)
Notes: These simulation results are based on the Adaptive Lasso algorithm, with penalization parameters chosen by BIC, under various true networks, network sizes and time periods
T=50, 100 and 150. In all cases, 1000 Monte Carlo iterations were performed. The true parameters are rho-0=.3, beta-0=.4 and gamma-0=.5. The % of true zeroes refers to the
proportion of true zero elements in the social interaction matrix that are estimated as smaller than .05. The % of true non-zeroes refers to the proportion of true elements greater than
.3 in the social interaction matrix that are estimated as non-zeros. The Mean Absolute Deviations are the mean absolute error of the estimated network compared to the true network
for the social interaction matrix W and the reduced form matrix respectively. The recovered parameter are the estimated parameters averaged across iterations. All specifications
include time and node fixed effects. Standard errors across iterations are in parentheses.
Table A5: Simulation Results, Adaptive Lasso
A. Erdos-Renyi B. Political party
N = 15 N = 30 N = 50 N = 15 N = 30 N = 50
ࢰ
ෝ
ࢼ
ෝࢽ
T=500 1000 1500 T=500 1000 1500 T=500 1000 1500 T=500 1000 1500 T=500 1000 1500 T=500 1000 1500
% True Zeroes .825 .878 .911 .884 .928 .958 .936 .966 .979 .824 .882 .916 .886 .932 .959 .940 .967 .979
(.018) (.020) (.021) (.007) (.007) (.006) (.019) (.003) (.002) (.020) (.021) (.020) (.007) (.007) (.005) (.003) (.003) (.002)
% True Non-Zeroes 1.000 1.000 1.000 1.000 1.000 1.000 .977 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000
(.000) (.000) (.000) (.000) (.000) (.000) (.107) (.000) (.000) (.000) (.000) (.000) (.000) (.000) (.000) (.000) (.000) (.000)
.039 .032 .028 .031 .025 .022 .025 .021 .019 .037 .030 .025 .030 .024 .021 .024 .020 .018
(.003) (.004) (.004) (.001) (.001) (.001) (.000) (.000) (.000) (.003) (.003) (.003) (.001) (.001) (.001) (.000) (.000) (.000)
.038 .027 .022 .039 .027 .022 .040 .027 .022 .038 .027 .022 .038 .027 .022 .039 .027 .022
(.002) (.001) (.001) (.001) (.001) (.001) (.001) (.000) (.000) (.002) (.001) (.001) (.001) (.001) (.001) (.001) (.000) (.000)
.488 .381 .362 1.000 .981 .617 1.000 1.000 1.000 .465 .396 .382 1.000 .953 .590 1.000 1.000 1.000
(.114) (.117) (.108) (.000) (.051) (.064) (.000) (.000) (.000) (.108) (.104) (.074) (.000) (.075) (.062) (.000) (.000) (.000)
.394 .398 .399 .343 .372 .400 .256 .350 .353 .397 .401 .399 .346 .382 .400 .280 .350 .350
(.017) (.009) (.006) (.017) (.025) (.003) (.064) (.002) (.013) (.022) (.014) (.005) (.013) (.024) (.000) (.029) (.002) (.000)
.964 .871 .809 1.000 1.000 .998 .956 1.000 1.000 .955 .841 .769 1.000 1.000 .995 1.000 1.000 1.000
(.046) (.079) (.083) (.000) (.000) (.009) (.205) (.000) (.000) (.053) (.068) (.046) (.000) (.000) (.015) (.000) (.000) (.000)
Notes: These simulation results are based on OLS estimates, under various true networks, network sizes and time periods T=500, 1000 and 1500. In all cases, 1000 Monte Carlo
iterations were performed. The true parameters are rho-0=.3, beta-0=.4 and gamma-0=.5. The % of true zeroes refers to the proportion of true zero elements in the social interaction
matrix that are estimated as smaller than .05. The % of true non-zeroes refers to the proportion of true elements greater than .3 in the social interaction matrix that are estimated as
non-zeros. The Mean Absolute Deviations are the mean absolute error of the estimated network compared to the true network for the social interaction matrix W and the reduced form
matrix respectively. The recovered parameter are the estimated parameters averaged across iterations. All specifications include time and node fixed effects. Standard errors across
iterations are in parentheses.
Table A6: Simulation Results, OLS
A. Erdos-Renyi B. Political party
N = 15 N = 30 N = 50 N = 15 N = 30 N = 50
ࢰ
ෝ
ࢼ
ෝࢽ
Table A7: Summary Statistics, Tax Competition Application
Obs Mean SD Min q25 Median q75 Max
State total tax per capita 1296 .371 .266 .036 .145 .300 .530 1.345
State income per capita 1296 9.951 2.130 4.105 8.585 9.919 11.375 18.808
Unemployment rate 1296 5.885 2.242 1.800 4.200 5.500 7.000 18.000
Proportion of young 1296 .234 .033 .160 .210 .240 .260 .310
Proportion of elderly 1296 .106 .020 .040 .090 .110 .120 .190
State governor's age 1296 51.088 7.441 33.000 45.000 50.000 56.000 73.000
Governor term limit dummy 1296 .258 .438 .000 .000 .000 1.000 1.000
State total tax per capita 2688 0.983 0.803 0.036 0.037 0.813 1.557 4.298
State income per capita 2736 13.268 4.016 4.147 10.348 12.960 15.879 27.974
Unemployment rate 2688 5.764 2.026 1.800 4.300 5.400 6.800 17.800
Proportion of young 2688 0.236 0.033 0.170 0.210 0.230 0.260 0.340
Proportion of elderly 2688 0.117 0.023 0.050 0.100 0.120 0.130 0.190
State governor's age 2736 53.557 8.134 33.000 47.000 53.000 59.000 78.000
Governor term limit dummy 2638 0.249 0.433 0.000 0.000 0.000 0.000 1.000
Notes: Summary statistics of variables (in levels) used in subsequent regressions. Besley and Case sample runs from 1962 to 1988 and
extended sample until 2014. State total tax per capita is the sum of sales, income and corporation tax in thousands of 1982 US dollars. Stateincome per capita in thousands of 1982 US dollars. Proportion of young is the proportion of the population between 5 and 17 years. Proportionof elderly is the proportion of the population aged 65 or older. State governor's age in years. Governor term limit dummy is equal to 1 if governorfaces term limits in the current mandate. Data sources: State total tax per capita, Census of Governments (1972, 1977, 1982, 1987, 1992-2016)and Annual Survey of Goverment Finances (all other years); Steta income per capita, Bureau of Economic Analysis; Unemployment rate,Bureau of Labor Statistics; Proportion of young (aged 5-17) and elderly (aged 65+), Census Population & Housing Data; State governor's ageand political variables manually sourced from individual governor's webpages on Wikipedia.
A. Besley and Case sample (1962-1988)
B. Extended sample (1962-2014)
Table A8: Exogenous Social Effects
Dependent variable: Change in per capital income and corporate taxes
Coefficient estimates, standard errors in parentheses
(1) Initial (2) OLS
(3) 2SLS: IVs are
Characteristics of
Neighbors
(4) 2SLS: IVs are
Characteristics of
Neighbors-of Neighbors
Economic Neighbors' tax change (t - [t-2]) .645 .145** .332* .608***
(.072) (.199) (.220)
Economic Neighbors' income per capita .090 .098*** .091*** .080***
(.011) (.012) (.014)
Economic Neighbors' unemployment rate 37.200 9.899*** 11.780*** 13.714***
(3.443) (2.856) (3.022)
Economic Neighbors' population aged 5-17 1378.1 376.2 478.5 596.6
(399.0) (414.2) (401.7)
Economic Neighbors' population aged 65+ -4304.5 -842.8 -769.7* -641.3
(504.3) (450.2) (468.6)
Economic Neighbors' governor age -2.158 -0.311 -0.293 -0.263
(.281) (.285) (.294)
Period
First Stage (F-stat) 9.571 10.480
Controls Yes Yes Yes Yes
State and Year Fixed Effects Yes Yes Yes Yes
Observations 2,952 2,952 2,544 2,592
1962-2015
Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The sample covers 48 mainland US states running from 1962 to 2015. The dependent
variable is the change in state i's total taxes per capita in year t. In all Columns, we penalize geographic neighbors in all Columns and allow forexogenous social effects. In OLS and IV regressions, the economic neighbors' effect is calculated as the weighted average of economic neighbors'variables. OLS regressions estimates are shown in Column 2. Column 3 shows the 2SLS regression where each geographic neighbors' tax change isinstrumented by lagged neighbor's state income per capita and unemployment rate. Column 4 shows a 2SLS regression where each geographicneighbors' tax change is instrumented by lagged neighbor-of-neighbor's state income per capita and unemployment rate. At the foot of Columns 3 and4 we report the p-value on the F-statistic from the first stage of the null hypothesis that instruments are jointly equal to zero. All regressions control forstate i’s income per capita in 1982 US dollars, state i’s unemployment rate, the proportion of young (aged 5-17) and elderly (aged 65+) in state i’spopulation, and the state governor's age. All specifications include state and time fixed effects. With the exception of governor's age, all variables aredifferenced between period t and period t-2. Robust standard errors are reported in parentheses.
Table A9: General Equilibrium Impacts of California Tax Rise
Geographic NeighborNetwork
Economic NeighborNetwork
Ratio
Average tax increase 0.0038 0.0066 1.74
Variance tax increase 0.0160 0.0153 0.96
Tax dispersion 0.0053 0.0141 2.66
States with tax increase 48 48 1.00
States with tax increase > 0.05% 11 44 4.00
States with tax increase > 0.5% 5 11 2.20
States with tax increase > 1% 4 8 2.00
States with tax increase > 2.5% 1 3 3.00
States with tax increase > 5% 1 1 1.00
Notes: This shows the equilibrium impulse responses in taxes set in each state as a result of California increasing its tax change by 10%. The rho
coefficient is derived from our preferred specification to estimate the economic network, where we penalize geographic neighbors to states, andallow for exogenous social effects (based on a sample of 48 mainland US states running from 1962 to 2015). We compare these derived taxchanges under the identified economic network structure, relative to that assumed under a geographic neighbors structure. The final Columnshows the ratio of the same statistic derived under each network.
A. % of zeros B. % of non-zeros
C. Mean Absolute Deviation of D. Mean Absolute Deviation of
E. Endogenous Social Effect, F. Exogenous Social Effect
Notes: These simulation results are based on the Adaptive Elastic Net GMM algorithm, with penalization parameters chosen
by BIC, under various true networks and time periods T=5, 10, 15, 25, 50, 100, 125 and 150. In all cases, 1000 Monte Carlo
iterations were performed. The true parameters are rho-0=.3, beta-0=.4 and gamma-0=.5. In Panel A, the % of zeroes refers
to the proportion of true zero elements in the social interaction matrix that are estimated as smaller than .05. In Panel B, the %
of non-zeros refers to the proportion of true elements greater than .3 in the social interaction matrix that are estimated as non-
zeros. In Panels C and D, the Mean Absolute Deviations are the mean absolute error of the estimated network compared to
the true network for the social interaction matrix W and the reduced form matrix respectively. In Panels E and F, the true
parameter values are marked in the horizontal red lines. The recovered parameter are the estimated parameters averaged
across iterations. All specifications include time and node fixed effects.
Figure A1: Simulation Results, Adaptive Elastic Net GMM
Π
ොߩ ොߛ
T=5 10 15 25 50 75 100 125 1500.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
T=5 10 15 25 50 75 100 125 1500.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
T=5 10 15 25 50 75 100 125 1500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
T=5 10 15 25 50 75 100 125 1500
0.2
0.4
0.6
0.8
1
1.2
T=5 10 15 25 50 75 100 125 1500
0.2
0.4
0.6
0.8
1
1.2
T=5 10 15 25 50 75 100 125 1500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Figure A2: Simulated and True Networks
Notes: These simulation results are based on the Elastic Net algorithm, with penalization parameters chosen by BIC, under various
true networks and time periods T=50, 100 and 150. In the two stylized networks (Erdos-Renyi and political party), we set N=30, and
the real world networks, the high school friendship and village network are based on N=65 and 70 non-isolated nodes respectively.
Party leaders in the political party network are marked in black in Panel B. In all cases, 1,000 Monte Carlo iterations were performed.
The true parameters are rho-0=.3, beta-0=.4 and gamma-0=.5. All specifications include time and node fixed effects. Kept edges are
depicted in blue: these links are estimated as non-zero in at least 5% of the iterations and are also non-zero in the true network.
Added edges are depicted in green: these links are estimated as non-zero in at least 5% of the iterations but the edge is zero in the
true network. Removed edges are depicted in red: these links are estimated as zero in at least 5% of the iterations but are non-zero in
the true network. The figures further distinguish between strong and weak links: strong links are shown in solid edges (whose strength
is greater than or equal to .3), and weak links are shown as dashed edges.
A. Erdos-Renyi B. Political Party
C. High-school D. Village
A. % of zeros B. % of non-zeros
C. Mean Absolute Deviation of D. Mean Absolute Deviation of
E. Endogenous Social Effect, F. Exogenous Social Effect
Figure A3: Simulation Results, Adaptive Elastic Net GMM
Erdos-Renyi graph (N=30)
Notes: Simulations with common shocks between the exogenous variable and the error term: time-constant and varying at the
individual level ("individual"), constant across individuals and varying over time ("time") and both types of shocks. These
simulation results are based on the Adaptive Elastic Net GMM algorithm, with penalization parameters chosen by BIC, under
various true networks and time periods T=25, 50, 100, 125 and 150. In all cases, 1000 Monte Carlo iterations were performed.
The true parameters are rho-0=.3, beta-0=.4 and gamma-0=.5. In Panel A, the % of zeroes refers to the proportion of true zero
elements in the social interaction matrix that are estimated as smaller than .05. In Panel B, the % of non-zeros refers to the
proportion of true elements greater than .3 in the social interaction matrix that are estimated as non-zeros. In Panels C and D,
the Mean Absolute Deviations are the mean absolute error of the estimated network compared to the true network for the social
interaction matrix W and the reduced form matrix respectively. In Panels E and F, the true parameter values are marked in the
horizontal red lines. The recovered parameter are the estimated parameters averaged across iterations. All specifications
include time and node fixed effects.
Alternative Structures of Shocks
Π
ොߩ ොߛ
T=25 50 75 100 125 1500.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
T=25 50 75 100 125 1500
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
T=25 50 75 100 125 1500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
T=25 50 75 100 125 1500.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
T=25 50 75 100 125 1500
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
T=25 50 75 100 125 1500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Positive values indicate higher equilibrium taxes under Economic neighbors than geographic neighborsNegative values indicate low equilibrium taxes under Economic neighbors than geographic neighbors
Figure A4: General Equilibrium Impacts of CA Tax Rise Shocks
State's Reaction to 10% increase in CA taxes
Log(equilibrium taxes under W-econ) - Log(equilibrium taxes under W-geo)
Notes: This shows the equilibrium impulse responses in taxes set in each state as a result of California increasing its tax change by 10%. This is as derived from our preferred
specification, where we penalize geographic neighbors to states, and allow for exogenous social effects. We compare these derived tax changes under the identified economic networkstructure, relative to that assumed under a geographic neighbors structure. We graph the log change in equilibrium taxes under economic neighbors, minus the log change in equilibriumtaxes under geographic neighbors. Positive values (red shaded) states indicate higher equilibrium taxes under economic neighbors than geographic neighbors, and negative values(blue shaded) states indicate lower equilibrium taxes under economic neighbors than geographic neighbors.